Heating ventilation air conditioning system fault detection method based on deep transfer learning and principal component analysis

Abstract

The invention discloses a heating ventilation air conditioning system fault detection method based on deep transfer learning and principal component analysis. Comprising the following steps: respectively processing a source domain data set and a target domain data set by using a principal component analysis method to obtain a source domain feature data set and a target domain feature data set; after a deep learning model is constructed, the deep learning model is pre-trained by using a source domain feature data set, parameters of a convolutional neural network module of the pre-trained model are frozen, and a trained deep learning model is obtained by using a target domain feature data set; and receiving and processing sensing signal data to be predicted by using the trained deep learning model to obtain a fault detection result of the heating ventilation air conditioning system. According to the method, dependence on a large amount of labeled data is reduced, fault information can be fed back in time, and fault detection is accurate and reliable.

Description

Technical Field

[0001] The present invention relates to a fault detection technology for a heating, ventilation and air conditioning system, and in particular to a fault detection method combining deep transfer learning and principal component analysis (PCA). Background Art

[0002] With the continuous advancement of building intelligence, the stability and efficiency of HVAC systems, as a key component of building internal environment control, are crucial to ensuring building comfort, energy saving and consumption reduction, and operational safety. Fault detection and diagnosis of HVAC systems is an important part of ensuring their normal operation.

[0003] However, existing fault detection technologies have some problems that need to be solved: Data dependency: Traditional fault detection methods, such as statistical-based models (ARIMA, VAR, etc.), rely on a large amount of historical data to build fault detection models. In the absence of sufficient data, the accuracy and generalization ability of these methods are limited.

[0004] Linearity assumption limitation: Many existing techniques assume that system behavior is linear or approximately linear, which limits their ability to capture the nonlinear characteristics and complex failure modes of the system.

[0005] Insufficient real-time performance: For systems that require real-time monitoring and rapid response, existing methods may not be able to provide fast enough fault detection speed, resulting in delayed fault response.

[0006] Multivariate data processing capabilities: HVAC systems generate a large amount of multivariate, multi-source time series data. Existing technologies often have difficulty effectively identifying and utilizing the complex relationships between variables when processing such data.

[0007] Generalization capability: Fault detection models that are customized for a single system or device usually lack generalization capabilities and are difficult to adapt to different systems or devices.

[0008] Explainability: Although deep learning methods perform well in fault detection, they are often regarded as "black boxes" and lack explainability, which is a significant disadvantage when the cause of the fault needs to be clarified. Summary of the invention

[0009] In view of the above problems, the present invention proposes a HVAC system fault detection method based on deep transfer learning and principal component analysis. The present invention uses transfer learning to solve the data dependency problem using a pre-trained model, uses principal component analysis (PCA) to improve the expressiveness and interpretability of features, and enhances the detection capability of complex fault modes through the powerful nonlinear mapping capability of the deep learning model. The method is also real-time and generalizable, and is applicable to various HVAC systems.

[0010] The technical solution adopted by the present invention is: 1. A Fault Detection Method for HVAC Systems Based on Deep Transfer Learning and Principal Component Analysis The HVAC system fault detection method comprises the following steps: S1) acquiring sensor signal data of the integrated automation system in normal operation and different fault states, and constructing them into source domain data sets; wherein the sensor signal data at least includes vibration signal data. The sensor signal data may also include one or more of the physical quantity signals that can be detected and converted by sensors, such as temperature signal data, humidity signal data, pressure signal data, indoor air quality data, flow data, water tank level data, electric valve opening, etc.

[0011] S2) Obtain sensor signal data of the HVAC system under normal operation and different fault conditions, and construct them as the target domain dataset.

[0012] S3) performing principal component analysis on the source domain dataset and the target domain dataset respectively to obtain a source domain feature dataset and a target domain feature dataset; In step S3, the principal component analysis method is specifically as follows: standardizing the data set and constructing a covariance matrix based on the standardized data set; performing eigendecomposition on the covariance matrix to obtain eigenvalues ​​corresponding to the covariance matrix and eigenvectors corresponding to each eigenvalue; according to the size of the eigenvalue, selecting the eigenvectors corresponding to the first few largest eigenvalues ​​as the principal component eigenvectors, the principal component eigenvectors constitute a new feature space, and projecting the standardized data set into the new feature space to obtain a feature data set.

[0013] S4) Build a deep learning model (Transfer-CNN-BiLSTM-Attention, TCBA); The deep learning model includes: The input layer is used to receive the input data of the deep learning model. The dimension of the input data is the total number of principal component feature vectors. The convolutional neural network (CNN) module is mainly composed of three convolutional layers and one pooling layer in series. The input of the first convolutional layer is connected to the output of the input layer. Batch Normalization (BN) layer, whose input is connected to the output of the pooling layer in the convolutional neural network module; A Bi-directional Long-Short Term Memory (BiLSTM) module, whose input is connected to the output of the batch normalization layer; The attention mechanism module has its input connected to the output of the bidirectional long short-term memory network module; Fully connected layer, the input is connected to the output of the attention mechanism module; A Softmax function layer (normalized exponential function layer), whose input end is connected to the output end of the fully connected layer, and the Softmax function layer converts a set of real numbers output by the fully connected layer into probability distributions corresponding to different fault categories; The output layer, whose input end is connected to the output end of the Softmax function layer, is used to output the probability distribution corresponding to different fault categories. The number of neurons in the output layer is the same as the number of target fault categories.

[0014] The convolutional neural network module is mainly composed of three convolutional layers and one pooling layer in series; in the three convolutional layers, the number of convolution kernels increases gradually, and the size of the convolution kernels is 2×1; the outputs of the three convolutional layers all use the maximum value of 2×1, and the pooling layer reduces the size of the output features by half.

[0015] In the three convolutional layers, the numbers of convolution kernels are 64, 128, and 256 respectively.

[0016] The Softmax function layer converts a set of real numbers output by the fully connected layer into probability distributions corresponding to different fault categories and outputs them through the output layer. The number of neurons in the output layer is the same as the number of target fault categories. The target fault categories include sensor faults, refrigeration system faults, fan faults, compressor faults, water system faults and other faults. In a specific implementation, the target fault category is related to the adopted data set.

[0017] S5) using the source domain feature data set to train the deep learning model to obtain a pre-trained model; freezing the parameters of the convolutional neural network module of the pre-trained model, and using the target domain feature data set to train the frozen pre-trained model to obtain a trained deep learning model; The training process of step S5 uses a cross entropy loss function to calculate the classification loss and updates the model parameters through back propagation; and uses the Adam optimizer as the optimization method for the model parameters.

[0018] The cross entropy loss function is set according to the following formula: L=—∑ M i=1 y i log(y i ’ ) Where M represents the number of target fault categories, y i represents the true probability value of the i-th target fault category, y i ’ Represents the probability of the i-th target fault category predicted by the deep learning model.

[0019] S6) Fault detection: Use the trained deep learning model to process the sensor signal data to be predicted to obtain the fault detection result. The process of step S6 is specifically as follows: S6.1) collecting sensor signal data of the target HVAC system as the sensor signal data to be predicted, and processing the sensor signal data to be predicted using a principal component analysis method to obtain feature data to be predicted; wherein the number of principal component feature vectors in the feature data to be predicted and the feature data set is the same; S6.2) inputting the feature data to be predicted into the input layer of the trained deep learning model, and the output layer of the trained deep learning model outputs the prediction result, which is the probability distribution corresponding to different fault categories; S6.3) According to the probability distribution corresponding to different fault categories, the final predicted fault detection result is obtained.

[0020] Step S6.3 is specifically as follows: if the maximum probability value in the probability distribution is greater than the preset probability threshold, the fault category corresponding to the maximum probability value is used as the final predicted fault detection result; if the maximum probability value in the probability distribution is less than or equal to the preset probability threshold, normal operation is used as the final predicted fault detection result.

[0021] 2. A computer device The system comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.

[0022] The beneficial effects of the present invention are: 1. The method of the present invention uses transfer learning to first use the source domain feature data set to train the deep learning model, and then uses the target domain feature data set to train the pre-trained model after freezing the first-layer model parameters, thereby reducing the dependence on a large amount of labeled data, so that effective fault detection can be performed even when data is limited.

[0023] 2. The method of the present invention not only achieves rapid response and real-time or near real-time fault detection through feature processing based on principal component analysis and an abnormal scoring system of a deep learning model, providing timely feedback for system operation and maintenance, but also combines the powerful feature extraction capability of deep learning and the dimensionality reduction and denoising effect of the principal component analysis method to improve the accuracy and reliability of fault detection.

[0024] 3. The method of the present invention improves the generalization ability of the method through the application of pre-trained deep learning models and principal component analysis methods, so that the present invention can adapt to different types of HVAC systems and various operating conditions.

[0025] 4. The method of the present invention is interpretable and helps operation and maintenance personnel to better understand fault characteristics and system status by analyzing principal components and feature changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the overall structure of the deep learning model in the present invention; Figure 2 Schematic diagram of a single long short-term memory network unit in the present invention; Figure 3 Schematic diagram of the attention mechanism module in the present invention; Figure 4 Schematic diagram of the transfer learning principle in the present invention; Figure 5 It is a schematic diagram of the fault identification process in the present invention. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] like Figure 5 As shown, the method of the present invention specifically comprises the following steps: S1) Construct source domain dataset: obtain sensor signal data of the integrated automation system under normal operation and different fault conditions, and construct it into a source domain dataset. The source domain dataset can provide rich data resources for the pre-training stage of the deep learning model.

[0029] The specific process of step S1 is: S1.1) Select data source: Select at least one integrated automation system as the data source.

[0030] Integrated automation systems include other equipment or systems that are similar to HVAC systems in terms of operating characteristics and possible failure types, but with richer data. These equipment and systems include, but are not limited to, industrial machinery, power equipment, and other types of HVAC (Heating, Ventilation and Air Conditioning) systems.

[0031] S1.2) Collect sensor signal data from different sensor sources under normal operation and different fault conditions of the integrated automation system selected in step S1.1 to ensure the diversity and comprehensiveness of the data set.

[0032] The sensor signal data includes one or more of the physical quantity signals that can be detected and converted by sensors, such as temperature signal data, humidity signal data, pressure signal data, indoor air quality data, flow data, water tank level data, electric valve opening, vibration signal data, etc. These sensor signal data are all time series data. It should be noted that the sensor signal data at least includes vibration signal data.

[0033] S1.3) Data preprocessing: The sensor signal data collected in step S1.2 is preprocessed. The preprocessing includes signal denoising, signal alignment, sampling rate unification and other steps to improve data quality and prepare for subsequent analysis and model training.

[0034] Signal denoising: Use an adaptive filter to perform wavelet denoising on the sensor signal data to reduce the noise component in the signal. The adaptive filter is set according to the following formula: s filtered (t) = D wavelet [s raw (t)] In the formula, s raw (t) represents the original sensor signal data, s filtered (t) represents the sensor signal data after filtering.

[0035] Signal alignment: For time delay or misalignment of signals, the signals can be aligned through cross-correlation. Signal alignment is performed using the following formula: s aligned (t) = s raw (t-τ) In the formula, s aligned (t) represents the sensor signal data after signal alignment, s raw (t-τ) represents the original sensor signal data, and τ represents the width of the signal misalignment.

[0036] Sampling rate unification: For signals with inconsistent sampling rates, resampling is required to unify the sampling rate. The sampling rate is unified using the following formula: s resampled (t)=resample(s raw (t),rate) In the formula, s resampled (t) represents the data after the sampling rate is unified, resample( ) represents the resampling function, s raw (t) represents the original sensor signal data, and rate represents the uniform sampling rate.

[0037] S1.4) Data labeling: Use the device status as a label to label the preprocessed signal data obtained in step S1.2, and combine all labeled signal data into a source domain dataset. The device status includes normal operation and different fault states, so as to facilitate supervised learning when training the deep learning model.

[0038] In a specific implementation, all sensor signal data of each device is taken as a sample, and each sample is labeled using the operating status of the device (normal operation or specific fault type).

[0039] S1.5) Dataset division: Divide the source domain dataset into training set, validation set, and test set. The training set is used for pre-training of the deep learning model, the validation set is used for parameter adjustment and optimization during the pre-training of the deep learning model, and the test set is used to evaluate the performance of the deep learning model after pre-training.

[0040] S2) Constructing a target domain dataset: Acquire sensor signal data of the HVAC system in normal operation and different fault states, and construct it as a target domain dataset.

[0041] In the process of constructing the target domain dataset, the key is to ensure that the target domain dataset can comprehensively cover all characteristics of the HVAC system in normal and abnormal states, especially the vibration characteristics.

[0042] The specific process of step S2 is: S2.1) Data collection planning: Determine the equipment and location for data collection, select key monitoring points that can represent the status of the HVAC system, and select the appropriate sensor type and quantity to ensure that the sensor signal data of the HVAC system can be captured, especially the vibration signal data.

[0043] S2.2) Data collection: When the HVAC system is operating normally, collect sensor signal data of a certain time series to ensure that the data covers different operating conditions and load conditions. Identify different types of failures that may occur in the HVAC system, artificially introduce or simulate these failures, and collect corresponding sensor signal data.

[0044] S2.3) Data preprocessing: The sensor signal data collected in step S2.2 is preprocessed. The preprocessing includes signal denoising, signal alignment, sampling rate unification and other steps to ensure data quality for subsequent feature extraction and model training.

[0045] S2.4) Data labeling: Using the device status as a label, label the preprocessed data obtained in step S2.3 to clearly distinguish between normal status and different types of fault status. All labeled signal data are combined into a target domain data set.

[0046] S2.5) Data segmentation: Split the target domain dataset into training set, validation set, and test set.

[0047] S3) Feature extraction and optimization: Use the principal component analysis method to process the source domain dataset and the target domain dataset respectively to obtain the source domain feature dataset and the target domain feature dataset. This step uses the principal component analysis method to achieve dimensionality reduction and denoising of the dataset, which can enhance the expressiveness of the features and reduce the complexity of the model.

[0048] In step S3), the principal component analysis method is specifically: ① Standardize the data set to obtain a standardized data set. Use the following formula for standardization: x'=(x-μ) / σ Among them, x represents the original eigenvalue in the data set, μ represents the mean of the feature, σ represents the standard deviation of the feature, and x' represents the standardized eigenvalue.

[0049] ② Construct a covariance matrix based on the standardized data set. The covariance matrix represents the linear relationship between the dimensions of the data. The element formula of the covariance matrix is ​​as follows: σ IJ =[1 / (n-1)]X T X=[1 / (n-1)]∑ n k=1 (x ’ ki x ’ kj ) Among them, σ IJ represents the covariance matrix, n represents the total number of samples, x ’ ki represents the i-th standardized eigenvalue of the k-th sample, x ’ kj represents the jth standardized eigenvalue of the kth sample, X represents the standardized matrix converted from the standardized data set, X T Represents the transpose of a normalized matrix.

[0050] ③ Perform eigendecomposition on the covariance matrix to obtain the eigenvalues ​​corresponding to the covariance matrix and the eigenvectors corresponding to each eigenvalue. The eigenvalues ​​represent the contribution of each eigenvector to the variance, and the eigenvectors represent the directions in the feature space.

[0051] ④ According to the size of the eigenvalue, select the eigenvectors corresponding to the first m largest eigenvalues ​​as the principal component eigenvectors. The number m of principal component eigenvectors is determined by the cumulative explained variance ratio, that is, the sum of the variance contributions of the selected principal component eigenvectors reaches a preset threshold (preset to 95% in the embodiment of the present invention). The specific process is: calculate the cumulative explained variance ratio of the principal component eigenvector according to the following formula: (∑ m i=1 λ i ) / (∑ p i=1 λ i ) Among them, m represents the number of principal component eigenvectors, p represents the total number of features, and λ i Represents the i-th eigenvalue corresponding to the covariance matrix.

[0052] The minimum m value with a cumulative explained variance ratio greater than a preset threshold is taken as the number of principal component eigenvectors.

[0053] ⑤Use the selected principal component eigenvectors as the basis. These principal component eigenvectors form a new feature space. Project the standardized data set into the new feature space to obtain a feature data set. Projection is performed according to the following formula: y=XW Among them, X represents the standardized matrix, W represents the matrix containing the principal component eigenvectors, and y represents the matrix form of the standardized principal component feature data set.

[0054] S4) Deep learning model design: Build a deep learning model (Transfer-CNN-BiLSTM-Attention, TCBA).

[0055] like Figure 1 As shown, the deep learning model includes: The input layer is used to receive the input data of the deep learning model. The dimension of the input data is the total number m of principal component feature vectors.

[0056] The convolutional neural network (CNN) module is mainly composed of three convolutional layers and one pooling layer in series. The input of the first convolutional layer is connected to the output of the input layer. The convolutional neural network module is mainly composed of three convolutional layers and one pooling layer in series. In each convolutional layer, the convolution kernel slides along the time axis to capture local patterns and trends. The convolution kernel formula is as follows: X t a =f(∑ n c=1 x c t-1 ×k t a +b t a ) In the formula, x t a represents the ath feature quantity output by the tth layer; f( ) represents the nonlinear activation function ReLU; k t a Represents the ath value of the convolutional layer vector of the tth layer; b t a represents the a-th bias of the t-th layer; In the three convolution layers, the number of convolution kernels increases gradually, and the size of the convolution kernel is 2×1; the output of the three convolution layers uses the maximum value of 2×1, and the pooling layer halves the size of the output feature to reduce the complexity of the network. In the three convolution layers, the number of convolution kernels is 64, 128, and 256 respectively, so as to obtain more advanced features at the high level of the model.

[0057] The Batch Normalization (BN) layer has its input connected to the output of the pooling layer in the convolutional neural network module.

[0058] The input of the Bi-directional Long-Short Term Memory (BiLSTM) module is connected to the output of the batch normalization layer. The Bi-directional Long-Short Term Memory module is set according to the following formula: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i )C ~ t =tanh(W C ·[h t-1 ,x t ]+b C ) o t =σ(W o ·[h t-1 ,x t]+b o )h t =o t tanh(C t ) C t =f t ·C t-1 +i t ·C ~ t Among them, f t 、i t , o t Respectively represent the activation values ​​of the forget gate, input gate, and output gate at the t-th time step, σ( ) represents the sigmoid function, and W f , W i , W o , W C、 b f , b i , b o and b C are learnable weights and biases. C t , C t-1 Represent the cell states at the current time step t and the previous time step t-1, respectively, C ~ t represents the candidate cell state, h t 、h t-1 denote the hidden states at time step t and the previous time step t-1 respectively.

[0059] like Figure 2 As shown in the figure, the bidirectional LSTM network module consists of two LSTM neural networks, one for processing the forward sequence and the other for processing the reverse sequence. Figure 1 In the figure, each rectangular box in the BiLSTM module represents a long short-term memory network LSTM unit: h t '=LSTM(h t-1 ',x t ) h t ''=LSTM(h t-1 '',x t ) y t =h t ' ⊕ h t '' Among them, h t ' represents the output of the forward LSTM at time step t, h t '' represents the output of the reverse LSTM at time step t, y tRepresents the output of the bidirectional long short-term memory network module after element-by-element addition.

[0060] The input of the attention mechanism module is connected to the output of the bidirectional long short-term memory network module. Figure 3 As shown, receiving the output y of the bidirectional long short-term memory network t As input, the corresponding attention score is calculated by the scoring function, and the corresponding weight is obtained after processing by the Softmax function. Finally, the sum of the product of the weight and the module input information is used as the output of the attention mechanism module. The attention mechanism module is set according to the following formula: u t =Vtanh(Wy t +b) a t =softmax(u t )=[exp(u t )] / ∑ i t=1 exp(u j ) c t =∑ i t=1 a t y t Among them, V and W represent the first weight matrix and the second weight matrix respectively, a t represents the feature vector weight, u j represents the summation term, u t represents the attention score, y t represents input information, b represents bias coefficient, c t Represents the output features.

[0061] The fully connected layer has its input connected to the output of the attention mechanism module. It is used to integrate and transform the output of the attention mechanism module for final classification. The output of the fully connected layer is a set of real numbers.

[0062] Softmax function layer (normalized exponential function layer), whose input is connected to the output of the fully connected layer; used to convert a set of real numbers output by the fully connected layer into a probability distribution; Output layer: The input end is connected to the output end of the Softmax function layer to output the prediction results of the deep learning model. The number of neurons in the output layer is the same as the number of target fault categories.

[0063] S5) Model training and migration: Figure 4As shown, according to the principle of transfer learning, the source domain feature dataset is used to train the deep learning model to obtain a pre-trained model; the parameters of the convolutional neural network module of the pre-trained model are frozen, and the frozen pre-trained model is trained using the target domain feature dataset to obtain a trained deep learning model.

[0064] In step S5), the training process uses the cross entropy loss function to calculate the classification loss and updates the model parameters through back propagation; the Adam optimizer is used as the optimization method for the model parameters.

[0065] Specifically, the cross entropy loss function is set according to the following formula: L=—∑ M i=1 y i log(y i ’ ) Where M represents the number of target fault categories, y i represents the true probability value of the i-th target fault category (usually 0 or 1, indicating whether the category is correct), y i ’ Represents the probability of the i-th target fault category predicted by the model.

[0066] Specifically, the Adam optimizer optimizes the model parameters according to the following formula: v t =β1v t-1 +(1-β1)·▽ θ L s t =β2s t-1 +(1-β2)·(▽ θ L) 2 θ t+1 =θ t -η / [s t / (1-β2 t ) 1 / 2 ]-v t where θ represents the model parameters (the aggregate gradient at time step t), L represents the loss function, and v t and t Represent the estimates of the first-order and second-order moments of the gradient, β1 and β2 represent the first and second decay rates, η represents the learning rate, ▽ θ represents the partial derivative of the loss function with respect to the model parameter θ, t represents the current time step, t-1 represents the previous time step, and t+1 represents the next time step.

[0067] Step S5) is specifically: S5.1) Input the training set and validation set in the source domain feature data set into the deep learning model for training. After each round of training, input the validation set into the model to monitor whether the deep learning model is continuously improving. When the preset training round is reached, that is, when the model converges, the model parameters are retained; S5.2) Input the test set in the source domain feature data set into the deep learning model saved in step S5.1 to evaluate the model performance. If it is not optimal, adjust the learning rate or training batch parameters and repeat step S5.1 until the model diagnosis accuracy reaches a preset threshold. At this time, it means that the performance of the deep learning model has reached the optimal level and it is used as a pre-training model for transfer learning. S5.3) selectively train the pre-trained model, i.e., freeze the parameters of the convolutional neural network module and fine-tune the parameters of the high-level network: use the training set and the validation set in the target domain feature dataset to train the pre-trained model. The training process is the same as step S5.1. After the training, the model parameters are retained; wherein the high-level network refers to the bidirectional long short-term memory network (BiLSTM) module, the attention mechanism module, and the fully connected layer; S5.4) Evaluate the pre-trained model saved in step S5.3 using the target domain test set, the steps are the same as step S5.2, and use the best pre-trained model as the trained deep learning model.

[0068] S6) Fault detection: Use the trained deep learning model to process the sensor signal data to be predicted to obtain the fault detection result.

[0069] The process of step S6) is specifically as follows: S6.1) using the principal component analysis method to process the sensor signal data to be predicted to obtain the feature data to be predicted; the number of principal component eigenvectors in the feature data to be predicted is consistent with the number of principal component eigenvectors in the feature data set obtained in step S3, that is, the number of principal component eigenvectors in the feature data to be predicted and the feature data set is the same, which is m; S6.2) Input the feature data to be predicted into the input layer of the trained deep learning model, and the feature data to be predicted is forward propagated through each layer of the trained deep learning model. Each layer calculates the output according to the trained parameters, and the output layer of the trained deep learning model outputs the probability distribution corresponding to different fault categories; S6.3) According to the probability distribution corresponding to different fault categories, the final predicted fault detection result is obtained.

[0070] Step S6.3 is specifically as follows: if the maximum probability value is greater than a preset probability threshold, the fault category with the maximum probability value is used as the final predicted fault detection result; otherwise, normal operation is used as the final predicted fault detection result.

[0071] The specific embodiments of the present invention are as follows: This embodiment includes the following steps: S1) Construct source domain dataset: Select sensor signals of other HVAC systems as source domain datasets. Collect sensor signal data of these devices in normal operation and different fault states. Preprocess the collected data, including denoising, signal alignment, sampling rate unification, etc. Divide the training set, validation set, and test set in a ratio of 4:2:4.

[0072] S2) Construction of target domain dataset: Select key monitoring points from the HVAC system and install sensors to collect vibration signals. Collect data of the system under normal operation and simulated fault conditions. Preprocess the data to ensure data quality. Divide the training set, validation set, and test set in a ratio of 4:2:4.

[0073] S3) Feature extraction: Use the principal component analysis method to process the source domain dataset and the target domain dataset respectively to obtain the source domain feature dataset and the target domain feature dataset.

[0074] S4) Build a deep learning model: Small convolution kernels of size 2×1 are used in the convolution layers to deepen the network and suppress overfitting, thereby improving the learning ability of the model. The output of the convolution module uses a 2×1 maximum pooling layer to halve the output feature size to reduce network complexity. The number of convolution kernels increases gradually in each layer, to 64, 128, and 256, respectively, so as to obtain more advanced features at the high level of the model.

[0075] S5) Model training and migration: In the pre-training stage, the model learning rate is set to 0.001, the optimizer is Adam, the training is repeated for 120 rounds, the batch size is 64, and the L2 regularization parameter is set to avoid overfitting of the model. The regularization parameter is set to 0.0001.

[0076] S6) Fault detection: Apply the trained deep learning model to the fault detection of the sensor signal data to be predicted, and realize rapid fault identification through the fault detection results. The output layer of the trained deep learning model outputs the prediction results, which are the probability distribution of different fault categories.

[0077] Use the Softmax function to convert the output into a probability for easy interpretation and decision making. Determine the most likely fault type based on the probability distribution of the model output. Set a probability threshold, and only when the predicted probability exceeds this threshold will it be considered a credible prediction result.

Claims

1. A HVAC system fault detection method based on deep transfer learning and principal component analysis, characterized in that: The following steps are involved: S1) acquiring sensor signal data of the integrated automation system and constructing it as a source domain data set; S2) acquiring sensor signal data of the HVAC system and constructing it as a target domain dataset; S3) performing principal component analysis on the source domain dataset and the target domain dataset respectively to obtain a source domain feature dataset and a target domain feature dataset; S4) constructing a deep learning model; the deep learning model includes a convolutional neural network module, a bidirectional long short-term memory network module and an attention mechanism module; S5) using the source domain feature data set to train the deep learning model to obtain a pre-trained model; freezing the convolutional neural network module, and using the target domain feature data set to train to obtain a trained deep learning model; S6) Use the trained deep learning model to process the sensor signal data to be predicted to obtain the fault detection result.

2. The HVAC system fault detection method based on deep transfer learning and principal component analysis according to claim 1, characterized in that: In step S1, sensors are used to collect sensor signal data of the integrated automation system in normal operation and in different fault states; in step S2, sensors are used to collect sensor signal data of the HVAC system in normal operation and in different fault states; The sensor signal data includes one or more of vibration signal data, temperature signal data, humidity signal data, pressure signal data, indoor air quality data, flow data, water tank level data and electric valve opening, and the sensor signal data includes at least vibration signal data.

3. The HVAC system fault detection method based on deep transfer learning and principal component analysis according to claim 1, characterized in that: In step S3, the principal component analysis method is specifically as follows: standardizing the data set and constructing a covariance matrix based on the standardized data set; performing eigendecomposition on the covariance matrix to obtain the eigenvalues ​​corresponding to the covariance matrix and the eigenvectors corresponding to each eigenvalue; according to the size of the eigenvalues, selecting the eigenvectors corresponding to the first m largest eigenvalues ​​as the principal component eigenvectors, the principal component eigenvectors constitute a new feature space, and projecting the standardized data set into the new feature space to obtain a feature data set.

4. The HVAC system fault detection method based on deep transfer learning and principal component analysis according to claim 3 is characterized in that: The minimum m value with a cumulative explained variance ratio greater than a preset threshold is taken as the number of principal component eigenvectors.

5. The HVAC system fault detection method based on deep transfer learning and principal component analysis according to claim 3 is characterized in that: The deep learning model includes: The input layer is used to receive the input data of the deep learning model. The dimension of the input data is the total number of principal component feature vectors. The convolutional neural network module is mainly composed of three convolutional layers and one pooling layer in series. The input end of the first convolutional layer is connected to the output end of the input layer. A batch normalization layer whose input is connected to the output of the pooling layer in the convolutional neural network module; A bidirectional long short-term memory network module, with its input connected to the output of the batch normalization layer; The attention mechanism module has its input connected to the output of the bidirectional long short-term memory network module; Fully connected layer, the input is connected to the output of the attention mechanism module; A Softmax function layer, whose input end is connected to the output end of the fully connected layer, and the Softmax function layer converts a set of real numbers output by the fully connected layer into probability distributions corresponding to different fault categories; The output layer, whose input end is connected to the output end of the Softmax function layer, is used to output the probability distribution corresponding to different fault categories to the deep learning model. The number of neurons in the output layer is the same as the number of fault categories.

6. The HVAC system fault detection method based on deep transfer learning and principal component analysis according to claim 5 is characterized in that: The convolutional neural network module is mainly composed of three convolutional layers and one pooling layer connected in series; in the three convolutional layers, the number of convolution kernels increases gradually in each layer.

7. The HVAC system fault detection method based on deep transfer learning and principal component analysis according to claim 1, characterized in that: In the training process of step S5, the classification loss is calculated using the cross entropy loss function and the model parameters are updated through back propagation, and the Adam optimizer is used as the optimization method for the model parameters; The cross entropy loss function is set according to the following formula: L=—∑ M i=1 and i log(y i ’ ) Where M represents the number of target fault categories, y i represents the true probability value of the i-th target fault category, y i ’ Represents the probability of the i-th target fault category predicted by the deep learning model.

8. The HVAC system fault detection method based on deep transfer learning and principal component analysis according to claim 1, characterized in that: The process of step S6 is specifically as follows: S6.1) collecting sensor signal data of the target HVAC system, processing it using a principal component analysis method, and obtaining feature data to be predicted; wherein the number of principal component feature vectors in the feature data to be predicted and the feature data set is the same; S6.2) Inputting the feature data to be predicted into the input layer of the trained deep learning model, the output layer of the trained deep learning model outputs the probability distribution corresponding to different fault categories; S6.3) According to the probability distribution corresponding to different fault categories, the final predicted fault detection result is obtained.

9. The HVAC system fault detection method based on deep transfer learning and principal component analysis according to claim 8, characterized in that: The step S6.3 is specifically as follows: when the maximum probability value in the probability distribution is greater than a preset probability threshold, the fault category corresponding to the maximum probability value is used as the final predicted fault detection result; otherwise, normal operation is used as the final predicted fault detection result.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 9 are implemented.

Images