HVAC fault detection method and system based on conditional drive and hybrid detection

Abstract

This invention discloses a fault detection method and system for HVAC systems based on condition-driven and hybrid detection. It selects three variables—indoor temperature, air volume, and duct pressure—of a normally operating HVAC system as indicator variables for condition-driven detection, converting time-slice data into condition-slice data. Feature extraction is performed on each condition slice using a spatiotemporal variational autoencoder (SVA) model to obtain a SVA similarity coefficient. Similarity analysis is then performed using the SVA similarity coefficient to cluster the condition slices into the same operating condition. A hybrid detection method using nonlinear and linear canonical variables is employed to calculate the standard control limits corresponding to each operating condition category. Online data is acquired, determining the operating condition category of the online measurement data and its online control limits. These online control limits are then compared with the standard control limits of the operating condition category to obtain a classification result indicating whether a fault has occurred in the online measurement data.

Description

Technical Field

[0001] This invention relates to the field of HVAC fault detection technology, and in particular to a fault detection method and system for HVAC based on condition-driven and hybrid detection. Background Technology

[0002] Heating, ventilation, and air conditioning (HVAC) systems refer to indoor systems or related equipment responsible for heating, ventilation, and air conditioning. HVAC systems control air temperature and humidity, improving indoor comfort and are a crucial component of medium to large-sized industrial and office buildings. During long-term operation, various malfunctions can occur due to component performance degradation and inadequate maintenance, such as sensor drift, stuck actuators like dampers, and pipe scaling. These malfunctions directly impact environmental comfort and increase energy consumption. Therefore, it is necessary to find a highly efficient and accurate method for detecting HVAC malfunctions.

[0003] In the actual operation of HVAC systems, influenced by various factors such as indoor temperature, air volume, duct pressure, and human control, HVAC systems operate under different steady-state conditions and switch between different operating states. Therefore, HVAC systems exhibit a wide range of non-stationary characteristics during use, encompassing both steady-state and transient states. For non-stationary processes, the frequent changes in process operating states cause the variance, mean, and dynamic relationships of process variables to change frequently over time. Furthermore, HVAC systems exhibit both linear and nonlinear characteristics. For example, linear characteristics exist between indoor temperature and time, and between airflow velocity and fan speed, while nonlinear characteristics exist between turbulent velocity and fan speed, and between the rate of heat radiation and the difference in indoor temperature. In HVAC systems, linear and nonlinear characteristics are often intertwined and mutually influential. Therefore, the monitoring, control, and optimization of the system require a comprehensive consideration of both linear and nonlinear characteristics. Existing HVAC fault detection methods generally do not consider these characteristics, resulting in the following problems:

[0004] (1) In actual operation, the changes in process operating conditions of HVAC systems will be affected by multiple indicator variables. Due to the frequent changes in operating conditions, these variables exhibit large-scale non-stationary characteristics. Existing single-mode or multi-mode research methods are time-driven. This traditional time axis analysis method often analyzes the change patterns over time and uses clustering methods to reveal the changes in process operating conditions. However, due to the frequent state switching of HVAC systems and the different target states of the switching, this brings great difficulties to modeling and understanding.

[0005] (2) When dividing the working conditions, it is necessary to extract the key features of the data for cluster analysis. However, existing clustering methods generally do not consider the correlation information and dynamic characteristics between various variables in the process data of HVAC system, and the data features are not fully extracted.

[0006] (3) Since linear and nonlinear characteristics are mixed in the actual operation of HVAC systems, a single nonlinear model is no longer the best choice. Therefore, hybrid linear-nonlinear modeling can provide a feasible alternative for mining process data features. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for HVAC fault detection based on condition-driven and hybrid detection.

[0008] On the one hand, a fault detection method for HVAC systems based on condition-driven and hybrid detection is provided, including:

[0009] (1) Select the three variables of indoor temperature, air volume and duct pressure of the normally operating HVAC system as the indicator variables of condition driving, and convert the time slice data into condition slice data.

[0010] (2) Feature extraction is performed on each condition slice using a spatiotemporal variational autoencoder model to obtain the spatiotemporal variational autoencoder similarity coefficient; then, similarity analysis is performed using the spatiotemporal variational autoencoder similarity coefficient to cluster the condition slices into the same working condition.

[0011] (3) The standard control limits corresponding to each type of working condition are calculated by using a mixed detection method of nonlinear canonical variables and linear canonical variables.

[0012] (4) Obtain online measurement data of the HVAC system, determine the operating condition class of the online measurement data in the same way as (1) to (2), determine the online control limit of the online measurement data in the same way as (3), compare the online control limit with the standard control limit of the operating condition class of the online measurement data, and obtain the classification result of whether the online measurement data has a fault.

[0013] On the other hand, a HVAC fault detection system based on condition-driven and hybrid detection is provided, including:

[0014] The conversion module is configured to select three variables—indoor temperature, air volume, and duct pressure—of a normally operating HVAC system as condition-driven indicator variables, and convert time-slice data into condition-slice data.

[0015] The feature extraction module is configured to: extract features on each condition slice using a spatiotemporal variational autoencoder model to obtain spatiotemporal variational autoencoder similarity coefficients; and then use the spatiotemporal variational autoencoder similarity coefficients to perform similarity analysis, clustering the condition slices into the same working condition.

[0016] The hybrid detection module is configured to use a hybrid detection method of nonlinear canonical variables and linear canonical variables to calculate the standard control limits corresponding to each type of working condition.

[0017] The classification output module is configured to: acquire online measurement data of the HVAC system; determine the operating condition class of the online measurement data using the same method as the transformation module and feature extraction module; determine the online control limits of the online measurement data using the same method as the hybrid detection module; compare the online control limits with the standard control limits of the operating condition class to which the online measurement data belongs; and obtain the classification result of whether the online measurement data has a fault.

[0018] Furthermore, an electronic device is also provided, including:

[0019] Memory, used for non-transitory storage of computer-readable instructions; and

[0020] Processor, for executing the computer-readable instructions,

[0021] When the computer-readable instructions are executed by the processor, they perform the method described in the first aspect above.

[0022] In another aspect, a storage medium is also provided for non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the instructions of the method described in the first aspect are executed.

[0023] In another aspect, a computer program product is also provided, including a computer program that, when run on one or more processors, is used to implement the method described in the first aspect above.

[0024] The above technical solution has the following advantages or beneficial effects:

[0025] To address the issue of widespread non-stationary characteristics in HVAC systems, we found that although operating conditions vary uncertainly over time, under the same operating conditions, process characteristics are largely similar and may follow certain relationships. That is, the process characteristics of the HVAC system change systematically along the conditional direction. Therefore, this invention employs a condition-driven method to reconstruct the three-dimensional data of the HVAC system. Since indoor temperature, air volume, and duct pressure are the most direct manifestations of the HVAC system's regulation process, these three variables are selected as indicator variables for condition-driven reconstruction. Time-slice data is transformed into condition-slice data, enabling the reconstruction of the HVAC system's three-dimensional time-varying data from the time axis to the conditional axis. This invention uses a multivariate condition-driven method to reconstruct the time-varying three-dimensional data of the HVAC system, selecting these variables as indicator variables. By reconstructing the time-varying three-dimensional data from the time axis to the conditional axis, the problem of widespread non-stationary operation of HVAC systems between and within batches can be effectively solved, significantly improving the accuracy of fault detection.

[0026] To address the correlation and dynamic characteristics of data across multiple condition slices and to fully extract process features and accurately classify multiple operating conditions, a spatiotemporal variational autoencoder (ST-VAE) is used for feature extraction on each condition slice of the HVAC system. ST-VAE reduces the dimensionality of the data to extract feature values ​​for the condition slices. Then, similarity analysis is performed using the ST-VAE similarity coefficient, clustering condition slices with high similarity into the same operating condition, thus achieving accurate classification of the multi-operating-condition model. The process features of condition slices within the same operating condition are similar, while those under different operating conditions are different. Establishing separate fault detection models for different operating conditions enables accurate detection of HVAC faults. This invention uses a method based on a ST-VAE to cluster condition slice data with high similarity. The encoder encodes the condition data into a latent space. To improve the interpretability and stability of the autoencoder, the latent vectors are modeled as Gaussian distributions, and the data is reconstructed using a decoder. The ST-VAE similarity coefficient is then used to calculate the similarity of the latent vectors, clustering condition slices with high similarity into the same operating condition.

[0027] To address the issue of mixed linear and nonlinear characteristics in the data of various operating conditions after the HVAC system is segmented, this invention employs a hybrid detection method combining Kernel Canonical Dissimilarity Analysis (KCVDA) and Canonical Dissimilarity Analysis (CVDA) to model the segmented operating conditions. First, KCVDA is applied to the collected process data to extract nonlinear canonical variables (NCVs) from the principal subspace of KCVDA. Then, CVDA is applied to the residual subspace of KCVDA to extract linear canonical variables (CVs) from the principal subspace of CVDA. Fault monitoring statistics T2 Q and are used to detect the residual matrices of the residual subspace of linear-nonlinear mixed canonical variables and CVDA, respectively. Since the prior distribution information of the nonlinear principal component features extracted from the normal data of each sub-operating condition is unknown, kernel density estimation is used to determine the control limits of the monitoring statistics. This invention combines kernel canonical dissimilarity analysis and canonical dissimilarity analysis algorithms to establish a fault detection model based on nonlinear-linear mixed monitoring for each operating condition, extracting mutually independent linear and nonlinear canonical variables from the datasets of each operating condition. This method can effectively handle the problem of mixed linear and nonlinear features, and has better accuracy compared to a single linear or nonlinear model. Attached Figure Description

[0028] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0029] Figure 1 This is a system structure diagram of Example 1;

[0030] Figure 2 This is a flowchart of the process for obtaining conditional slices in Example 1;

[0031] Figure 3 This is a structural diagram of the spatiotemporal variational autoencoder in Example 1;

[0032] Figure 4 This is a schematic diagram of the statistical modeling of hybrid detection in Example 1. Detailed Implementation

[0033] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0034] Example 1

[0035] This embodiment provides a fault detection method for HVAC systems based on condition-driven and hybrid detection;

[0036] like Figure 1 As shown, the HVAC fault detection method based on condition-driven and hybrid detection includes:

[0037] S101: Select the indoor temperature, air volume, and duct pressure of a normally operating HVAC system as indicator variables for condition-driven data, and convert the time slice data into condition slice data.

[0038] S102: Feature extraction is performed on each condition slice using a spatiotemporal variational autoencoder model to obtain the spatiotemporal variational autoencoder similarity coefficient; then, similarity analysis is performed using the spatiotemporal variational autoencoder similarity coefficient to cluster the condition slices into the same working condition;

[0039] S103: The standard control limits for each type of working condition are calculated by using a mixed detection method of nonlinear canonical variables and linear canonical variables.

[0040] S104: Obtain online measurement data of the HVAC system, determine the operating condition class of the online measurement data in the same way as S101 to S102, determine the online control limits of the online measurement data in the same way as S103, compare the online control limits with the standard control limits of the operating condition class of the online measurement data, and obtain the classification result of whether the online measurement data has a fault.

[0041] Further, in step S101: the indoor temperature, air volume, and duct pressure of a normally operating HVAC system are selected as condition-driven indicator variables, and the time-slice data is converted into condition-slice data, specifically including:

[0042] S101-1: In the time-varying three-dimensional normal operation dataset X(I×K×J) of the HVAC system, perform a slicing operation perpendicular to the time axis to obtain two-dimensional time slices X at different times. k (I×J); where k represents the sampling time, k=1,2,…,K, I represents the batch, and J represents the variable;

[0043] S101-2: The k-th two-dimensional time slice X k In the variable J of (I×J), calculate the mean value of the indoor temperature variable for different batches at the same time. Average air volume Mean of pipeline pressure Indicator variables, where k = 1, 2, ..., K, are defined by the following formula:

[0044]

[0045] Among them, T k (I) represents the indoor temperature value of the i-th batch at the k-th time, V k (I) represents the air supply volume value of the i-th batch at the k-th time, P k (I) represents the duct pressure value of the I-th batch at k time points.

[0046] To better achieve the division of operating conditions, multivariate variables are transformed into weighted variables. These weighted variables combine multiple indicator variables through a weighting method. Firstly, After standardization, the weights of the three variables are calculated to achieve a composite representation of multiple indicator variables. The weighted variable formula is as follows:

[0047]

[0048] S101-3: Slice the weighted variable conditional slices according to the increasing weighted variable [x] min ,x min+ ,…,x i ,x max Arranged in the order of ], the weighted variable x is obtained on the condition axis. k Conditional slices X that change sequentially from smallest to largest m (I×J); where X m (I×J), m=1,2,…,K represents the m-th condition slice.

[0049] Furthermore, after S101-2 and before S101-3, the following steps are also included:

[0050] S101-23: The chimpanzee optimization algorithm is used to optimize the three parameters α, β, and γ in the formula:

[0051] S101-231: Chimpanzees randomly select their initial location;

[0052] S101-232: Position Update Steps: The position is updated using an update position formula, with different parameters selected. The update position formula is as follows:

[0053]

[0054] in, It represents the position of individual i at time step t. It is the coefficient of the individual with the lowest objective function value. R1 and R2 are the positions of another individual randomly selected from the population. R1 and R2 are random numbers, which usually follow a normal distribution with a mean of 0 and a standard deviation of 1.

[0055] S101-233: Weight Update Steps: To find the optimal weights, an objective function is introduced to update the weights: Among them, y k Represents the actual value.

[0056] S101-233: Repeat the position update step and weight update step until the predetermined number of iterations is reached or the stopping condition is met.

[0057] It should be understood that manually selecting the three weighting parameters does not necessarily yield the optimal solution. To achieve optimal algorithm performance, a discrete chimpanzee optimization algorithm based on the nearest neighbor traction operator is used to optimize the three parameters α, β, and γ in the formula. This method simulates the sociological behavior and learning mechanisms within a chimpanzee group, combined with the nearest neighbor traction operator, to achieve adaptive parameter adjustment. The goal is to minimize the weighting variable x. k The difference from the actual situation is the optimization objective. Through iterative iteration, the weight parameters α, β, and γ are finally optimized.

[0058] It should be understood that the aforementioned Heating, Ventilation, and Air Conditioning (HVAC) system refers to a system used to provide and regulate the comfort of the interior environment of a building, integrating the three major functions of heating, ventilation, and air conditioning to ensure that the temperature, humidity, and air quality of the indoor environment meet the requirements of human comfort. Heating systems raise indoor temperatures by heating air or water, ventilation systems improve air quality by introducing fresh air and expelling stale air, and air conditioning systems lower indoor temperatures by cooling or dehumidifying air. An efficient HVAC system not only provides a comfortable indoor environment but also achieves energy conservation and emission reduction, and lowers operating costs. Its objectives include temperature control, humidity control, air quality, and air circulation.

[0059] It should be understood that the time-varying three-dimensional normal operation dataset includes: Time-varying: The data changes over time, meaning the dataset contains multiple data samples collected at different time points. The time-varying dataset reflects the dynamic changes of the system or process over time and can capture trends and patterns that change over time. Three-dimensional: The dataset has three dimensions, which can be understood as each data sample having three different attributes at a certain point in time. In this invention, these three attributes are time K, batch I, and variable J.

[0060] It should be understood that the two-dimensional time slice includes: in the time-varying three-dimensional dataset, each time point K contains two attributes, batch and variable, so this time point K is a two-dimensional data, and therefore time point K is regarded as a two-dimensional time slice of the time-varying three-dimensional dataset.

[0061] It should be understood that, considering the continuous operation of an HVAC system over several days, if the operation status of a workday is defined as an operation batch I, then each operation batch contains two-dimensional data of time K and variable J. Arranging the normal operating data of different operation batches together forms a time-varying three-dimensional dataset of the HVAC system. The operation data of different batches over several consecutive days exhibit time-varying dynamic characteristics at the batch level; and within each day, the data within a batch also exhibits a wide range of non-stationary operation characteristics over time.

[0062] To address the issue of time-varying dynamic characteristics in the batch-based operation data of HVAC systems over multiple days, and the significant non-stationary operation within each day's data, a multivariate condition-driven method is employed to reconstruct the three-dimensional time-varying normal operation data of the HVAC system from the time axis to the condition axis. This effectively handles the non-stationary operation problem of frequent switching between multiple operating conditions in actual HVAC system operation. The data processing flow is as follows: Figure 2 As shown.

[0063] It should be understood that the time-varying three-dimensional dataset of the HVAC system is reconstructed from the time axis to the condition axis, resulting in condition slices arranged in ascending order of weighted variables.

[0064] Furthermore, such as Figure 3 As shown, step S102 involves: extracting features from each condition slice to obtain a similarity coefficient; then using the similarity coefficient to perform similarity analysis, clustering the condition slices into the same working condition, specifically including:

[0065] Each conditional slice is standardized, and the trained spatiotemporal variational autoencoder model is used to extract features from each standardized conditional slice to obtain the spatiotemporal variational autoencoder similarity coefficient.

[0066] The trained spatiotemporal variational autoencoder model includes: an encoder, a hidden layer, and a decoder connected in sequence; the encoder includes: a first convolutional layer, a temporal convolutional network (TCN), and a second convolutional layer connected in sequence; the decoder includes: a third convolutional layer, a graph convolutional network (GCN), and a fourth convolutional layer connected in sequence.

[0067] Similarity analysis is performed using spatiotemporal variational autoencoder similarity coefficients to cluster condition slices into the same working condition.

[0068] Furthermore, the standardization process for each condition slice includes:

[0069] conditional slice X m (I×J) is standardized to obtain a standard condition slice, and the formula is as follows:

[0070]

[0071] in, This represents the m-th standard condition slice. Indicates condition slice X m The mean of D(X) m ) represents condition slice X m The variance.

[0072] Furthermore, the trained spatiotemporal variational autoencoder model includes:

[0073] The encoder will process the standardized condition slices. The conditional stride is achieved through 1×1 convolutional embeddings, and multi-layer one-dimensional unrolled convolutions are used to progressively learn batch dependencies and compress data length for batches. Using a filter f(i) with kernel size K and unwinding factor d, perform a 1×k unwinding convolution operation, which is defined as follows:

[0074]

[0075] Where t(s) represents the conditional step size s of the compressed sequence t. The encoder compresses the original data of length T into a hidden state of length 1 through multiple one-dimensional unwinding convolutions, and then outputs the hidden state h of the intrinsic pattern through 1×1 convolution.

[0076] Based on the idea of ​​variational inference, the hidden layer is used to fit μ and σ, as shown in the following equation:

[0077] μ=TCN μ (X) (3)

[0078] logσ 2 =TCN σ (X) (4)

[0079] Where μ is the mean vector matrix and σ is the variance matrix; TCN μ (X) is responsible for extracting features from the input data X and generating a mean vector matrix μ, TCN σ (X) Generate a log-variance matrix value logσ for calculating the distribution of the latent variables. 2 .

[0080] The decoder, for the hidden state z(s), performs a transpose unwinding convolution operation using a filter f(i) with a kernel size of 1×k and an unwinding factor of d, which is defined as follows:

[0081]

[0082] Furthermore, the training process of the trained spatiotemporal variational autoencoder model includes:

[0083] Construct a training set, which is time-varying three-dimensional data of a heating, ventilation, and air conditioning system under normal operating conditions;

[0084] The spatiotemporal variational autoencoder model is trained using a training set. Training is stopped when the loss function value of the spatiotemporal variational autoencoder model no longer decreases, and the trained spatiotemporal variational autoencoder model is obtained.

[0085] The original data X is encoded into the hidden layer, and then the hidden layer is decoded to reconstruct the data. To minimize the reconstruction error of the spatiotemporal variational autoencoder model, the loss function is designed as follows:

[0086]

[0087] in, Let KL be the mean squared error loss function, and KL[q(·)||p(·)] be the Kullback-Leibler divergence between q(·) and p(·). The Kullback-Leibler divergence is an index used to measure the difference between two probability distributions.

[0088] Further, feature extraction is performed on each standardized conditional slice using the trained spatiotemporal variational autoencoder model to obtain spatiotemporal variational autoencoder similarity coefficients, including:

[0089] The loss function L is minimized by training the spatiotemporal variational autoencoder model. Then, the latent vector of the hidden layer is output from the trained spatiotemporal variational autoencoder model.

[0090] Z=q(X|N(μ,σ 2 (9)

[0091] The latent vectors of the hidden layer are spatiotemporal variational autoencoder similarity coefficients.

[0092] Furthermore, the step of using spatiotemporal variational autoencoder similarity coefficients for similarity analysis to cluster condition slices into the same working condition includes:

[0093] After outputting the latent vector matrix, the m-th conditional slice X is calculated using the spatiotemporal variational autoencoder similarity coefficient. m and the (m+1)th conditional slice X m+1 The similarity ρ between two conditional slices is calculated using the following formula:

[0094]

[0095] Where z m z m+1 They are conditional slices X m X m+1 The latent vector of z; θ is the latent vector of z. m z m+1 The angle between them;

[0096] If the similarity coefficient between adjacent multivariate condition slices is 0.5≤ρ≤1, it indicates that the feature similarity between the two multivariate condition slices is high, and the two multivariate condition slices are clustered into the same condition.

[0097] If the spatiotemporal variational autoencode similarity coefficient between adjacent multivariate condition slices is 0 ≤ ρ ≤ 0.5, it indicates that the feature similarity between two adjacent multivariate condition slices is low. The m-th multivariate condition slice is then taken as the last multivariate condition slice of the previous working condition, and the (m+1)-th multivariate condition slice is taken as the first multivariate condition slice of the next working condition. The spatiotemporal variational autoencoder model is used to accurately divide the Z sub-working condition models.

[0098] It should be understood that the spatiotemporal variational autoencoder similarity coefficient is actually the mean of the cosine of the angle between the latent vectors of the two types of data, and should be within the interval [0,1]. The similarity coefficient between conditional slices of the same type will be close to 1, which is greater than the similarity coefficient between conditional slices of different types.

[0099] It should be understood that existing clustering models generally do not consider the correlation information and dynamic characteristics between various variables in HVAC system data, leading to inaccurate segmentation of multiple operating conditions in HVAC systems. This invention employs a multi-condition segmentation module based on a spatiotemporal variational autoencoder to segment condition slices X. m (I×J) is used for feature extraction, and then the spatiotemporal variational autoencoder similarity coefficient is used to cluster the condition slices with high similarity into the same working condition, so as to obtain multiple different working condition data with large differences and realize the accurate division of multi-working condition model.

[0100] It should be understood that, firstly, feature values ​​are extracted by autoencoder within the divided condition slices, and then similarity analysis is performed using the spatiotemporal variational autoencoder similarity coefficient. Condition slices with high correlation are clustered into the same working condition, thereby achieving accurate division of multi-working-condition datasets based on condition slices.

[0101] Furthermore, S103: A hybrid detection method using nonlinear canonical variables and linear canonical variables is employed to calculate the standard control limits corresponding to each type of operating condition, including:

[0102] For each type of working condition, a corresponding fault detection model based on nonlinear-linear hybrid monitoring is established;

[0103] S103-1: A fault detection model based on nonlinear-linear hybrid monitoring performs kernel canonical dissimilarity analysis on the process data of the current operating condition dataset: extracting nonlinear canonical variables from the principal subspace of kernel canonical dissimilarity analysis; applying canonical dissimilarity analysis to the residual subspace of kernel canonical dissimilarity analysis; and extracting linear canonical variables and residual variables from the principal subspace of canonical dissimilarity analysis.

[0104] S103-2: Calculate the first statistic based on nonlinear and linear canonical variables;

[0105] S103-3: Calculate the second statistic based on the residual variable;

[0106] S103-4: Based on the first and second statistics, the kernel density estimation algorithm is used to calculate the standard control limits corresponding to each type of operating condition.

[0107] Further, S103-1: Based on the fault detection model of nonlinear-linear hybrid monitoring, kernel canonical dissimilarity analysis is performed on the process data of the current operating condition dataset: nonlinear canonical variables are extracted from the principal subspace of the kernel canonical dissimilarity analysis; canonical dissimilarity analysis is applied to the residual subspace of the kernel canonical dissimilarity analysis; linear canonical variables and residual variables are extracted from the principal subspace of the canonical dissimilarity analysis, specifically including:

[0108] set up For the process input at time k, For the process output at time k, define a past data vector containing past inputs and outputs. and future data vectors containing future outputs

[0109]

[0110]

[0111] Where p and f are the past and future data vectors z, respectively. p (k) and y f (k) time lag.

[0112] Suppose a training dataset has N measurements u(k) and y(k), k = 1, 2, ..., N. Under normal operating conditions, N is collected. For all k ∈ [p+1, p+M], z... p (k) and y f (k) Construct the Hankel matrix Z for the past and future. p and Y f ,as follows:

[0113]

[0114] Kernel Canonical Dissimilarity Analysis (KCVDA) uses kernel functions to map the original data to a high-dimensional feature space. This mapping captures nonlinear relationships into the features, so KCVDA is used to extract nonlinear features.

[0115] Assuming past input vector Nonlinear transformation Φ1: and future input vector Nonlinear transformation Φ2: Zp ,Y f It is a data sequence matrix containing information about the past and future, H i This is called the feature space.

[0116] Then, the goal is to find f1∈H1 and f2∈H2 such that f1(Z) p )=<Φ1(Z p ),f1> and f2(Y f )=<Φ2(Y f f1 and f2 have the highest correlation. f1 and f2 lie in linear spaces S1 and S2 respectively, and are spanned by the images of Φ1 and Φ2. Therefore:

[0117]

[0118] Using kernel matrix K p ,K f Mapping the original data to a high-dimensional feature space, K p ,K f The definition is as follows:

[0119]

[0120] Kernel canonical dissimilarity analysis (KCVDA) uses canonical dissimilarity (KCVD) between canonical variables projected in the past and future, with the KCVDd at time k being... s (k) is defined as follows:

[0121] d s (k)=AK f -BK p (14)

[0122] Where A∈(β1,β2,...,β) s It consists of the first s singular values, B∈(β) s+1 ,β s+2 ,...,β M It consists of the following s to m singular values.

[0123] Canonical Variable Dissimilarity Analysis (CVDA) is applied to the residual subspace of Kernel Canonical Variable Dissimilarity Analysis (KCVDA). CVDA focuses on finding the most discriminative linear projection directions between datasets, thus extracting linear features. CVDA utilizes canonical dissimilarity (CVD) between canonical variables in the residual subspace. The CVD at time k is... n (k) is defined as follows:

[0124] d n (k)=Ln y f (k)-S n J n z p (k) (15)

[0125] in, U n S contains the first n columns of U. n Composed of n maximum singular values ​​S n =diag(λ1,λ2,...,λ) n )composition.

[0126] It should be understood that process input data refers to the input information received by a system or process. This data can be any information that affects the system's behavior or output results, and is typically provided when the system or process begins execution.

[0127] It should be understood that process output data refers to the results or outputs generated by a system or process after its execution. This data typically reflects the system's or process's response to or processing of input data. Process output data can be any information or signal generated by the system, such as measurements, simulation results, predicted values, state variables, etc. This data reflects the processed input data and is used to analyze system behavior, evaluate performance, or make further decisions.

[0128] It should be understood that the future output future data vector refers to the set of f output data that the HVAC system will generate in the future, which are predicted forward.

[0129] It should be understood that a Hankel matrix is ​​a square matrix in which the elements on each antidiagonal are identical. A Hankel matrix is ​​a symmetric matrix where each element a... ij It depends only on the sum of i+j.

[0130] It should be understood that Kernel Canonical Variate Dissimilarity Analysis (KCVDA) is a statistical technique that combines kernel methods and canonical analysis to identify and separate different classes or groups in data.

[0131] It should be understood that canonical dissimilarity is the determination of the differences between linear data.

[0132] Further, S103-2: Based on nonlinear canonical variables and linear canonical variables, calculate the first statistic, including:

[0133] Calculate HotellingT at time k 2 Statistical Fault Detection Quantity T 2 :

[0134] T 2 (k)=d(k)d(k) T (16)

[0135] Among them, T 2 The change in the state vector d(k) is measured, where d(k) includes the nonlinear canonical dissimilarity d. s (k) and the dissimilarity of linear canonical variables d n (k).

[0136] Further, S103-3: Calculating the second statistic based on the residual variable includes:

[0137] Assume y r This is data in the CVDA residual space, y r The calculation formula is as follows:

[0138]

[0139] Introducing statistic Q:

[0140] Q(k)=y r (k)y r (k) T (18)

[0141] Q(k) measures the change in the linear residual space.

[0142] It should be understood that HotellingT 2 A t-statistic is a statistical method that extends the univariate t-test, used for comparing means in multivariate data scenarios. It can be viewed as a distance measure in a multivariate variable space, measuring the degree of difference between the sample mean vector and the hypothesized mean vector.

[0143] Furthermore, such as Figure 4 As shown, S103-4: Based on the first and second statistics, the kernel density estimation algorithm is used to calculate the standard control limits corresponding to each type of operating condition, including:

[0144] Since the prior distribution information of the mixed canonical variable characteristics is unknown, kernel density estimation is used to construct T based on normal operating condition data. 2 Control limits for the Q statistic and Q statistic.

[0145] Let the density δ(·) be unknown. For the mixed canonical variable features g1, g2, ... g N T 2 Statistic. T 2 The kernel density estimate of the statistic is:

[0146]

[0147] Where K(·) is a non-negative function with an integral of 1 and a mean of zero, h is the bandwidth parameter, and N is the number of samples for the sub-condition;

[0148] Let the density η(·) be unknown for Q1, Q2, ..., Q N y r (1),y r (2),…,y r The SPE statistic of (N) is given by the kernel density estimate of the SPE statistic:

[0149]

[0150] Understandably, normal data from different sub-operating conditions are input into the fault detection model. First, KCVDA is performed to extract NCVs. Then, CVDA is applied to extract linear features from the CVs in the residual subspace of KCVDA, and monitoring statistics T are constructed accordingly. 2 and Q.

[0151] It should be understood that the SPE statistic, short for Squared Prediction Error, is a commonly used statistic in process monitoring and fault detection, used to measure the difference between observed values ​​and model predictions.

[0152] Due to the mixed linear and nonlinear characteristics in the actual operation of HVAC systems, a single nonlinear model is no longer the optimal choice. To address the issue of mixed linear and nonlinear characteristics in the data of various operating conditions after the HVAC system is divided, this invention employs a hybrid detection method of KCVDA-CVDA to model each of the divided sub-operating conditions.

[0153] Further, S104: compares the online control limits with the standard control limits of the operating condition class to which the online measurement data belongs, and obtains a classification result of whether the online measurement data has a fault, including:

[0154] Statistics on online test data like This indicates that the test dataset X of the HVAC system test It's normal; otherwise, it indicates a malfunction.

[0155] The statistical measure Q of online test data test ,like This indicates that the test dataset X of the HVAC system test It's normal; otherwise, it indicates a malfunction.

[0156] Example 2

[0157] This embodiment provides a HVAC fault detection system based on condition-driven and hybrid detection.

[0158] A condition-driven and hybrid detection-based HVAC fault detection system includes:

[0159] The conversion module is configured to select three variables—indoor temperature, air volume, and duct pressure—of a normally operating HVAC system as condition-driven indicator variables, and convert time-slice data into condition-slice data.

[0160] The feature extraction module is configured to: extract features on each condition slice using a spatiotemporal variational autoencoder model to obtain spatiotemporal variational autoencoder similarity coefficients; and then use the spatiotemporal variational autoencoder similarity coefficients to perform similarity analysis, clustering the condition slices into the same working condition.

[0161] The hybrid detection module is configured to use a hybrid detection method of nonlinear canonical variables and linear canonical variables to calculate the standard control limits corresponding to each type of working condition.

[0162] The classification output module is configured to: acquire online measurement data of the HVAC system; determine the operating condition class of the online measurement data using the same method as the transformation module and feature extraction module; determine the online control limits of the online measurement data using the same method as the hybrid detection module; compare the online control limits with the standard control limits of the operating condition class to which the online measurement data belongs; and obtain the classification result of whether the online measurement data has a fault.

[0163] It should be noted that the aforementioned conversion module, feature extraction module, hybrid detection module, and classification output module correspond to steps S101 to S104 in Embodiment 1. The examples and application scenarios implemented by these modules and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should also be noted that these modules, as part of the system, can be executed in a computer system, such as a set of computer-executable instructions.

[0164] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0165] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0166] Example 3

[0167] This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0168] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0169] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0170] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0171] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0172] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0173] Example 4

[0174] This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0175] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault detection method for HVAC systems based on condition-driven and hybrid detection, characterized in that, include: (1) Select the indoor temperature, air volume, and duct pressure of a normally operating HVAC system as indicator variables for condition-driven data, and convert the time-slice data into condition-slice data, specifically including: Time-varying three-dimensional normal operation dataset of HVAC system In this process, slicing is performed perpendicular to the time axis to obtain two-dimensional time slices at different times. ;in, Indicates the sampling time. , Indicates batch, Represents variables; In the A two-dimensional time slice variables In the middle, the mean value of indoor temperature variables for different batches at the same time is calculated. Average air volume Average value of pipeline pressure Indicator variables, where, The formula is as follows: (1) in, This represents the indoor temperature value of the I-th batch at time k. This represents the air supply volume value of the i-th batch at time k. This represents the duct pressure value of the i-th batch at k time points; To better achieve the division of operating conditions, multivariate variables are transformed into weighted variables. These weighted variables combine multiple indicator variables through a weighting method. Firstly, , , After standardization, the weights of the three variables are calculated to achieve a composite representation of multiple indicator variables. The weighted variable formula is as follows: (2) Slice the weighted variables into conditional slices in ascending order of weighted variables. The order of the variables is used to obtain weighted variables on the condition axis. Condition slices that change sequentially from smallest to largest ;in, , Indicates the first A conditional slice; (2) Feature extraction is performed on each condition slice using a spatiotemporal variational autoencoder model to obtain the spatiotemporal variational autoencoder similarity coefficient; then, similarity analysis is performed using the spatiotemporal variational autoencoder similarity coefficient to cluster the condition slices into the same working condition. (3) Using a hybrid detection method combining nonlinear and linear canonical variables, the standard control limits corresponding to each type of working condition are calculated, including: For each type of working condition, a corresponding fault detection model based on nonlinear-linear hybrid monitoring is established; Based on a fault detection model using a nonlinear-linear hybrid monitoring system, kernel canonical dissimilarity analysis (KDA) is performed on the process data of the current operating condition dataset. This involves extracting nonlinear canonical variables from the principal subspace of the KDA; applying canonical dissimilarity analysis to the residual subspace of the KDA; and extracting linear canonical variables and residual variables from the principal subspace of the KDA. Specifically, this includes: set up For a moment Process input, For a moment The process output defines a past data vector containing past inputs and outputs. and future data vectors containing future outputs. ; ； ； in and These are past and future data vectors, respectively. and Time lag; Suppose a training dataset has Measured values and , Collected under normal operating conditions For all ,Depend on and Constructing the Hankel matrix of the past and future and ,as follows: (11) Kernel Canonical Dissimilarity Analysis (KCVDA) uses kernel functions to map the original data to a high-dimensional feature space. This mapping captures nonlinear relationships into the features, so KCVDA is used to extract nonlinear features. Assuming past input vector nonlinear transformation and future input vector nonlinear transformation ,in , It is a data sequence matrix containing information about the past and future. This is called the feature space; Then, the goal is to find and , making and It has the highest correlation; and Located in linear space and In the middle, they are made by , The image is stretched, therefore: (12) Using kernel matrix , Map the original data to a high-dimensional feature space. , The definition is as follows: (13) Kernel canonical dissimilarity analysis (KCVDA) utilizes canonical dissimilarity between canonical variables projected in the past and future, at time [time value missing]. KCVD The definition is as follows: (14) in, It consists of the first s singular values. It consists of the following s to m singular values; Canonical Variable Dissimilarity Analysis (CVDA) is applied to the residual subspace of Kernel Canonical Variable Dissimilarity Analysis (KCVDA). CVDA focuses on finding the most discriminative linear projection directions between datasets, thus extracting linear features. CVDA utilizes canonical dissimilarity (CVD) between canonical variables in the residual subspace. (Time step...) CVD The definition is as follows: (15) in, , Include The former List, Depend on The largest singular value composition; Calculate the first statistic based on nonlinear and linear canonical variables; Calculate the second statistic based on the residual variable; Based on the first and second statistics, the kernel density estimation algorithm is used to calculate the standard control limits corresponding to each type of working condition. (4) Obtain online measurement data of the HVAC system, determine the operating condition class of the online measurement data in the same way as (1)~(2), determine the online control limit of the online measurement data in the same way as (3), compare the online control limit with the standard control limit of the operating condition class of the online measurement data, and obtain the classification result of whether the online measurement data has a fault.

2. The HVAC fault detection method based on condition-driven and hybrid detection as described in claim 1, characterized in that, Feature extraction was performed on each condition slice using a spatiotemporal variational autoencoder model to obtain the spatiotemporal variational autoencoder similarity coefficient; Then, similarity analysis is performed using the spatiotemporal variational autoencoder similarity coefficient to cluster the condition slices into the same working condition, specifically including: Each conditional slice is standardized, and the trained spatiotemporal variational autoencoder model is used to extract features from each standardized conditional slice to obtain the spatiotemporal variational autoencoder similarity coefficient. The trained spatiotemporal variational autoencoder model includes: an encoder, a hidden layer, and a decoder connected in sequence; the encoder includes: a first convolutional layer, a temporal convolutional network, and a second convolutional layer connected in sequence; the decoder includes: a third convolutional layer, a graph convolutional neural network, and a fourth convolutional layer connected in sequence. Similarity analysis is performed using spatiotemporal variational autoencoder similarity coefficients to cluster condition slices into the same working condition.

3. The HVAC fault detection method based on condition-driven and hybrid detection as described in claim 2, characterized in that, The trained spatiotemporal variational autoencoder model includes: The encoder will process the standardized condition slices. Conditional step size through Convolutional embeddings are used, and multiple layers of one-dimensional unwinding convolutions are employed to progressively learn batch dependencies and compress data length for each batch. Using a kernel size of K and an expansion factor of Filter Perform unrolling convolution The operation is defined in the following form: (2-1) (2-2) in, Represents compressed sequence conditional step size The encoder uses multiple one-dimensional unwinding convolutions to convert the length of the input into a single input. The original data is compressed into a hidden state of length 1, and then... Convolution outputs the hidden state of the intrinsic pattern ; Based on the idea of ​​variational inference, the hidden layer is used to fit µ and σ, as shown in the following equation: (3) (4) in It is a mean vector matrix. It is the variance matrix; It is responsible for extracting features from the input data X and generating a mean vector matrix. , Generate a log-variance matrix value for calculating the distribution of the latent variables. ; Decoder, for hidden state Using a kernel size of The expansion factor is Filter The transpose unwind convolution operation is defined as follows: (6)； Feature extraction is performed on each standardized conditional slice using a trained spatiotemporal variational autoencoder model to obtain spatiotemporal variational autoencoder similarity coefficients, including: By using the trained spatiotemporal variational autoencoder model, the loss function is made The minimum value is obtained; then the latent vector of the hidden layer is output through the trained spatiotemporal variational autoencoder model: (9) The latent vectors of the hidden layer are spatiotemporal variational autoencoder similarity coefficients.

4. The HVAC fault detection method based on condition-driven and hybrid detection as described in claim 1, characterized in that, The method of using spatiotemporal variational autoencoder similarity coefficients for similarity analysis to cluster condition slices into the same working condition includes: After outputting the latent vector matrix, the similarity coefficient of the spatiotemporal variational autoencoder is calculated. m Conditional slices and the m +1 conditional slice Similarity between two conditional slices The formula is as follows: (10) In the formula, , These are condition slices , The latent vector; yes , The angle between them; If the similarity coefficient of adjacent multivariate conditional slices This indicates that the feature similarity between the two multivariate condition slices is high, and the two multivariate condition slices are clustered into the same condition. If the spatiotemporal variational autoencode similarity coefficient of adjacent multivariate conditional slices If the feature similarity between two adjacent multivariate condition slices is low, the m-th multivariate condition slice is taken as the last multivariate condition slice of the previous working condition, and the (m+1)-th multivariate condition slice is taken as the first multivariate condition slice of the next working condition.

5. The HVAC fault detection method based on condition-driven and hybrid detection as described in claim 1, characterized in that, Based on nonlinear and linear canonical variables, the first statistic is calculated, including: In the Hotelling time calculation Statistical Fault Detection Quantity : (16) in, Measurement state vector change, Including nonlinear canonical dissimilarity Dissimilarity with linear canonical variables ; The second statistic is calculated based on the residual variable, including: Assumption It is data in the CVDA residual space. The calculation formula is as follows: (17) Introducing statistics : (18) in, Measure the changes in the linear residual space.

6. The HVAC fault detection method based on condition-driven and hybrid detection as described in claim 1, characterized in that, Based on the first and second statistics, the kernel density estimation algorithm is used to calculate the standard control limits for each type of operating condition, including: Since the prior distribution information of the mixed canonical variable features is unknown, kernel density estimation is used to construct a model based on normal operating condition data. and Control limits for statistics; Let density Unknown For mixed canonical variable characteristics of Statistic; The kernel density estimate of the statistic is: （19） in It is a non-negative function with an integral of 1 and a mean of zero. It's a bandwidth parameter. The number of samples for each sub-condition; Let density Unknown for of Statistic, The kernel density estimate of the statistic is: （20）。 7. A HVAC fault detection system based on condition-driven and hybrid detection, characterized in that, include: The conversion module is configured to select three variables—indoor temperature, air volume, and duct pressure—of a normally operating HVAC system as condition-driven indicator variables, and convert time-slice data into condition-slice data. Specifically, this includes: Time-varying three-dimensional normal operation dataset of HVAC system In this process, slicing is performed perpendicular to the time axis to obtain two-dimensional time slices at different times. ;in, Indicates the sampling time. , Indicates batch, Represents variables; In the A two-dimensional time slice variables In the middle, the mean value of indoor temperature variables for different batches at the same time is calculated. Average air volume Average value of pipeline pressure Indicator variables, where, The formula is as follows: (1) in, This represents the indoor temperature value of the I-th batch at time k. This represents the air supply volume value of the i-th batch at time k. This represents the duct pressure value of the i-th batch at k time points; To better achieve the division of operating conditions, multivariate variables are transformed into weighted variables. These weighted variables combine multiple indicator variables through a weighting method. Firstly, , , After standardization, the weights of the three variables are calculated to achieve a composite representation of multiple indicator variables. The weighted variable formula is as follows: (2) Slice the weighted variables into conditional slices in ascending order of weighted variables. The order of the variables is used to obtain weighted variables on the condition axis. Condition slices that change sequentially from smallest to largest ;in, , Indicates the first A conditional slice; The feature extraction module is configured to: extract features on each condition slice using a spatiotemporal variational autoencoder model to obtain spatiotemporal variational autoencoder similarity coefficients; and then use the spatiotemporal variational autoencoder similarity coefficients to perform similarity analysis, clustering the condition slices into the same working condition. The hybrid detection module is configured to use a hybrid detection method combining nonlinear and linear canonical variables to calculate the standard control limits for each type of operating condition, including: For each type of working condition, a corresponding fault detection model based on nonlinear-linear hybrid monitoring is established; Based on a fault detection model using a nonlinear-linear hybrid monitoring system, kernel canonical dissimilarity analysis (KDA) is performed on the process data of the current operating condition dataset. This involves extracting nonlinear canonical variables from the principal subspace of the KDA; applying canonical dissimilarity analysis to the residual subspace of the KDA; and extracting linear canonical variables and residual variables from the principal subspace of the KDA. Specifically, this includes: set up For a moment Process input, For a moment The process output defines a past data vector containing past inputs and outputs. and future data vectors containing future outputs. ; ； ； in and These are past and future data vectors, respectively. and Time lag; Suppose a training dataset has Measured values and , Collected under normal operating conditions For all ,Depend on and Constructing the Hankel matrix of the past and future and ,as follows: (11) Kernel Canonical Dissimilarity Analysis (KCVDA) uses kernel functions to map the original data to a high-dimensional feature space. This mapping captures nonlinear relationships into the features, so KCVDA is used to extract nonlinear features. Assuming past input vector nonlinear transformation and future input vector nonlinear transformation ,in , It is a data sequence matrix containing information about the past and future. This is called the feature space; Then, the goal is to find and , making and It has the highest correlation; and Located in linear space and In the middle, they are made by , The image is stretched, therefore: (12) Using kernel matrix , Map the original data to a high-dimensional feature space. , The definition is as follows: (13) Kernel canonical dissimilarity analysis (KCVDA) utilizes canonical dissimilarity between canonical variables projected in the past and future, at time [time value missing]. KCVD The definition is as follows: (14) in, It consists of the first s singular values. It consists of the following s to m singular values; Canonical Variable Dissimilarity Analysis (CVDA) is applied to the residual subspace of Kernel Canonical Variable Dissimilarity Analysis (KCVDA). CVDA focuses on finding the most discriminative linear projection directions between datasets, thus extracting linear features. CVDA utilizes canonical dissimilarity (CVD) between canonical variables in the residual subspace. (Time step...) CVD The definition is as follows: (15) in, , Include The former List, Depend on The largest singular value composition; Calculate the first statistic based on nonlinear and linear canonical variables; Calculate the second statistic based on the residual variable; Based on the first and second statistics, the kernel density estimation algorithm is used to calculate the standard control limits corresponding to each type of working condition. The classification output module is configured to: acquire online measurement data of the HVAC system; determine the operating condition class of the online measurement data using the same method as the transformation module and feature extraction module; determine the online control limits of the online measurement data using the same method as the hybrid detection module; compare the online control limits with the standard control limits of the operating condition class to which the online measurement data belongs; and obtain the classification result of whether the online measurement data has a fault.

Images