Heating ventilation and air conditioning system fault detection and diagnosis method based on depth double-layer self-coding

Abstract

The invention discloses a heating ventilation air-conditioning system fault detection and diagnosis method based on depth double-layer self-coding, and relates to the technical field of heating ventilation air-conditioning system fault monitoring. According to the method, the deep auto-encoder is used for carrying out data dimension reduction, learning the rule of fault-free data, designing a DAE-C fault detection model, diagnosing the fault state of the monitoring data, and then designing a DAE-BIGRU fault diagnosis model to determine the fault type, so that the acquisition of the fault data and a label thereof is not needed, and the fault detection efficiency is improved. And calculation loss caused by analysis of a large amount of fault-free data in the fault diagnosis process is reduced, and the problem that label data of the HVAC system is lacked is solved, so that faults of the HVAC system can be detected and diagnosed even under the condition that no clear fault label exists.

Description

Technical Field

[0001] The present application relates to the technical field of HVAC system fault monitoring, and in particular to a HVAC system fault detection and diagnosis method based on deep double-layer autoencoding. Background Art

[0002] Energy consumption of heating, ventilation and air conditioning (HVAC) systems (HVAC systems for short) is particularly prominent in the building industry. HVAC systems often have some operational failures due to insufficient maintenance, improper installation or equipment failure, and fail to achieve the expected energy savings. These failures include mechanical failures, such as dampers or actuators; valves may also have some problems, such as stuck, damaged, leaking or freezing; sensor failures include sensor hardware failures and sensor drift failures, etc. These failures can lead to excessive energy consumption and energy waste. Therefore, fault detection and diagnosis (FDD) technology is a key technology to ensure building performance and improve energy efficiency. One obstacle to the application of big data analysis in buildings is the lack of semantic metadata standardization. Current FDD methods are mainly divided into data-driven or knowledge-driven. For fault detection, most researchers use data-driven methods including classification-based, unsupervised learning-based and regression-based methods. Knowledge-driven methods include model-based and rule-based methods. Supervised learning technology has demonstrated extremely high diagnostic accuracy and excellent automation in HVAC system fault detection and diagnosis tasks, and is the first choice in the "fault label is known" scenario. The label of the data sample indicates whether the sample belongs to the normal class or the fault class, and which fault class the sample belongs to. Supervised technology can use labeled data samples to train fault classifiers.

[0003] Although fault data is crucial for fault detection and diagnosis of HVAC systems, in real scenarios, the probability of HVAC system failure is much lower than the probability of fault-free operation, so obtaining sufficient and accurate fault data samples faces great challenges, and labeling fault data is often expensive. After a fault occurs, the system usually quickly repairs itself. If the data monitoring is not sensitive enough or the fault data monitoring is not comprehensive enough, the number of available fault data samples will be further reduced. In addition, manually collected fault data may have problems such as mislabeling, missing labels, and inconsistent labeling, which become obstacles to the application of supervised learning FDD technology. Unsupervised fault detection and diagnosis eliminates the dependence on labeled training data, so it has received widespread attention in recent years. The starting point of this type of method is that when an HVAC system fails, it will show abnormal patterns that are different from the normal working state. Through the inherent data patterns learned in the normal fault-free operation data, the unsupervised algorithm can identify and label any atypical states that may occur in the equipment, that is, detect and diagnose system faults.

[0004] At present, in view of the sparsity of fault data in HVAC systems, many researchers have tried FDD methods based on deep learning. For example, some studies have used Transformer encoders for semi-supervised learning methods for HVAC fault detection, adopted strategic data masking with Markov chain methods, eliminated the need for labeled data, and used the peak threshold method to determine faults dynamically and data-driven by abnormal thresholds. Another study proposed a method that combines the use of neural networks (NN) and mixture probabilistic principal component analyzers (MPPCA), which aims to fit complex data distributions through mixed models, so that the model can adapt to and describe complex data distributions well, that is, those instances that are significantly different from the known data distribution are effectively identified as outliers.

[0005] There are few studies on fault detection and diagnosis under unsupervised data in HVAC systems. A study proposed a method to discover the functional feed relationship between air handling units (AHU) and variable-air-volume (VAV) terminal units from sensor data. The study created a large temperature perturbation point to capture the relationship between AHU and VAV. The heat transfer process of VAV is the key basis for the study. A study studied the relationship between air conditioning units and variable air volume air conditioning terminals by establishing the relationship equation between the damper position data of variable air volume air conditioning terminals and the air supply duct pressure of air conditioning units. This statistical method is too dependent on the existing logical relationship of specific data. Other studies have achieved automatic functional relationship inference of AHU and VAV terminal units by detecting common events reflected in the readout data. The study used Markov event model, a posteriori reasoning, state-dependent transformation and other methods to identify the original time series and event series, and then derive the true functional relationship. This method is suitable for situations where there is less sensor data and events occur simultaneously, and has strong adaptability. However, sensor time series are independently modeled event detection, and this method sometimes misses the relationship between entities.

[0006] In summary, most current methods rely on human expertise, data integrity, data labels, or additional physical perturbation intervention, which will bring risks in labor costs and uncertainties in data labeling. Summary of the invention

[0007] The purpose of this application is to provide a HVAC system fault detection and diagnosis method based on deep two-layer autoencoding to solve the problem of lack of label data for HVAC systems, so that HVAC system faults can be detected and diagnosed even in the absence of clear fault labels.

[0008] To achieve the above objectives, this application provides the following solutions.

[0009] In a first aspect, the present application provides a HVAC system fault detection and diagnosis method based on deep double-layer autoencoding, comprising:

[0010] Construct a fault-free dataset and an unknown dataset for HVAC systems.

[0011] The first deep autoencoder in the DAE-C fault detection model is trained based on the fault-free data set to obtain a trained DAE-C fault detection model.

[0012] Based on the trained DAE-C fault detection model, fault detection is performed on the monitoring data in the unknown data set to obtain fault reconstruction data.

[0013] Based on the reconstruction data of each fault, the fault feature linkage analysis strategy is used to generate time window fault reconstruction data and its fault type pseudo label.

[0014] According to the fault reconstruction data in each time window and its fault type pseudo-label, the DAE-BIGRU fault diagnosis model is optimized and trained to obtain the trained DAE-BIGRU fault diagnosis model.

[0015] Based on the trained DAE-BIGRU fault diagnosis model, fault diagnosis of the HVAC system is performed.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects.

[0017] The present application provides a HVAC system fault detection and diagnosis method based on deep two-layer autoencoder. The present application uses deep autoencoder to perform data dimensionality reduction, learns the rules of fault-free data, designs DAE-C fault detection model, diagnoses the fault status of monitoring data, and then designs DAE-BIGRU fault diagnosis model to determine the fault type. There is no need to obtain fault data and its labels, and the computational loss caused by the fault diagnosis process for analyzing a large amount of fault-free data is reduced, which solves the problem of lack of label data for HVAC systems, so that even in the absence of clear fault labels, HVAC system faults can be detected and diagnosed. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 A flowchart of a method for fault detection and diagnosis of a HVAC system based on deep double-layer autoencoding is provided in accordance with an embodiment of the present application.

[0020] Figure 2 A schematic diagram of a HVAC system fault detection and diagnosis method based on deep double-layer autoencoding provided in an embodiment of the present application.

[0021] Figure 3 This is a general schematic diagram of a building provided in accordance with an embodiment of the present application.

[0022] Figure 4 A schematic diagram of the SDAHU and connection area provided in one embodiment of the present application.

[0023] Figure 5This is an exemplary graph of SDAHU time series data provided in one embodiment of the present application.

[0024] Figure 6 A structural schematic diagram of a DAE-C fault detection model provided in an embodiment of the present application.

[0025] Figure 7 A structural schematic diagram of the DAE-BIGRU fault diagnosis model provided in one embodiment of the present application.

[0026] Figure 8 A schematic diagram of component symptom chain inference for different fault types provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0028] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0029] In an exemplary embodiment, Figure 1 and Figure 2 As shown, a HVAC system fault detection and diagnosis method based on deep double-layer autoencoding is provided, including the following steps 101 to 106.

[0030] Step 101: construct a fault-free data set and an unknown data set of a HVAC system.

[0031] Step 102: training a first deep autoencoder in a DAE-C fault detection model based on the fault-free data set to obtain a trained DAE-C fault detection model.

[0032] Step 103 , performing fault detection on the monitoring data in the unknown data set based on the trained DAE-C fault detection model to obtain fault reconstruction data.

[0033] Step 104 : Based on each fault reconstruction data, a fault feature linkage analysis strategy is used to generate time window fault reconstruction data and its fault type pseudo label.

[0034] Step 105 , optimizing and training the DAE-BIGRU fault diagnosis model according to the fault reconstruction data of each time window and its fault type pseudo-label, to obtain a trained DAE-BIGRU fault diagnosis model.

[0035] Step 106: perform fault diagnosis on the HVAC system based on the trained DAE-BIGRU fault diagnosis model.

[0036] Implement the above steps 101 to 106. First, use the deep autoencoder to reduce the data dimension, learn the rules of fault-free data, design the DAE-C fault detection model, and diagnose the fault state of the sample data. Then design the DAE-BIGRU model based on BIGRU and multi-head attention mechanism (i.e., DAE-BIGRU fault diagnosis model) to determine the fault type. This method does not require the acquisition of fault data and its labels, and reduces the computational loss caused by the analysis of a large amount of fault-free data in the fault diagnosis process. It solves the problem of lack of label data for HVAC systems, so that even without clear fault labels, HVAC system faults can be detected and diagnosed.

[0037] At present, no research has attempted to completely infer the fault chain of HVAC from the beginning of the fault to the detection and generation of the fault. Therefore, it is necessary to detect and diagnose complex fault modes. While realizing fault diagnosis, the fault feature linkage analysis strategy is used to analyze the cause of the fault, which effectively supplements the lack of systematic evaluation of the impact of SDAHU measurement in existing research.

[0038] After completing fault detection and diagnosis in HVAC systems, it is very important to perform fault causal reasoning. Fault causal reasoning not only helps to determine the component or root cause that caused the fault, but also helps to determine the system fault propagation path, so it is of great significance for HVAC maintenance and prevention.

[0039] In another exemplary embodiment, the above method further includes the following steps 107 and 108.

[0040] Step 107 , based on the fault type of the fault reconstruction data in each time window, a reward function optimization decision-making method is used to determine the fault symptom chain of the HVAC system.

[0041] Step 108 , based on the fault symptom chain of the HVAC system, using the mapping relationship between features and components, determine the order in which each component exhibits abnormal behavior during the fault process, and construct a component symptom chain diagram for each fault type.

[0042] In another exemplary embodiment, a HVAC system including a single-duct VAV system is used to provide heating and cooling to, for example, the middle floors of an office building. The conditioned floor space consists of a single interior zone and four perimeter zones. The AHU distributes conditioned air to each zone through five VAV boxes. The hot water used by the AHU heating coil and the VAV box reheat coil is provided by a natural gas boiler. The cold water used by the AHU cooling coil is provided by a central chiller. Figure 3 Shows the overall layout of the building.

[0043] The simulated AHU used in the embodiment of the present application is as follows: Figure 4 As shown. The main components of AHU include supply fans with variable frequency drives (Vacuum Fluorescent Display, VFD), return fans with variable frequency drive, cooling and heating coils, cooling and heating control valves, and outdoor air and return air dampers. This variable air volume box is a typical single-pipe pressure-independent variable air volume box with reheating. The main components of the variable air volume box are reheat coils, reheat coil valves and terminal dampers.

[0044] Taking HVAC as an example, the measurement values ​​of the single-duct HVAC system (single-duct VAV-AHU) are shown in Table 1. BAS (Building Automation System) collects system operation interval data, electricity meter data, and environmental data (i.e., outdoor air temperature and relative humidity). Data is collected through various sensors installed on different subsystems and equipment of the HVAC system or through control signals sent by BAS or equipment. Among them, "BAS roll call" is an existing label for BAS data used in the experiment. The type indicates whether the data is sensor data or control data. In the embodiment of the present application, BAS collected and stored data points for one year, including a total of 115,000 points. A total of 1 fault-free data set and 5 fault data sets were collected (in the verification process, the fault data set and the fault-free data set constitute an unknown data set for algorithm verification. In the actual working process, the unknown data set contains monitoring data of unknown faults and fault types).

[0045] Table 1 Measurement values ​​of SDAHU system

[0046]

[0047]

[0048] In the BAS, monitoring data is recorded in two ways. In the first way, time series data is recorded in a continuous manner, that is, time series data is continuously sampled and recorded. In the tested BAS, supply air temperature data is collected and stored every 1 minute.

[0049] In another exemplary embodiment, the above step 101 may be replaced by the following steps 201 to 203 .

[0050] Step 201, obtain the monitoring data of the SDAHU in Table 1 every 1 minute, a total of 30 data per second, and collect a fault-free data set D no_fault = {X1, X2, ..., XP} and 5 different types of fault data sets D fault1 , D fault2 , …, D fault5 .

[0051] Step 202: Fuse the fault-free data set with the five fault data sets. First, add non-overlapping time windows to each data set, use the data label of the last time step in the window as the label of the window, and select all data sets from the fault-free data set. window =D no_fault [t last ], Select 20% of the data in the continuous time period from the faulty data set and the data read at the same time in the fault-free data set to merge into an unknown data set According to the working principle diagram, different features are mapped to different components. Because the embodiment of the present application needs to verify the subsequent method, the embodiment of the present application adopts an unknown data set obtained by the fusion of a fault-free data set and a faulty data set. In the actual working process, the unknown data set contains monitoring data of unknown faults and fault types. In the introduction of the subsequent method process, the fault-free data and faulty data in the unknown data set can also be regarded as monitoring data of unknown faults and fault types.

[0052] Step 203: Use a steady-state filter to remove dynamic data D from the data obtained in step 202 filtered = F(Draw) and after removing outliers and standardizing D normalized =σD filtered -μ.

[0053] After obtaining the BAS monitoring big data, the system operation data changes greatly during system start-up and shutdown, load mutation and operation mode switching. Such dynamic data is of no value for learning fault detection models. A steady-state filter is used to remove dynamic data. Several variables are selected as the input of the steady-state filter. Use K = X max -X min / X meanCalculate the slope K of each variable. If the sum of the slopes of all inputs exceeds the threshold, the system is in a transient state. The monitoring data in the current moving window is considered to be worthless and can be deleted. An example of a one-day time series of SDAHU is shown in Figure 5 As shown in the figure, it is obvious that these data series show a lot of differences. For example, some data are smooth, some data never seem to change, some data fluctuate regularly, some data have obvious peaks, and some data appear to be random.

[0054] In another exemplary embodiment, time window is a crucial concept in time series analysis. It can capture data patterns within a specific time period and is particularly useful for detecting and diagnosing HVAC system failures.

[0055] In the fault detection process, non-overlapping windows are used, that is, each time window is independent and there is no intersection between adjacent windows. The purpose of non-overlapping windows is to accurately model data patterns and simplify the calculation and training process of the model. In fault diagnosis, overlapping time windows are used. The overlapping window mechanism allows the window to slide on the data with a certain step size, extracting the data in each time window as a sample, containing the feature data of all time steps in the window. Overlapping windows can capture data behavior changes more carefully and perform fault diagnosis more accurately through changes in data behavior patterns.

[0056] For each time window, the data label of the last time step in the window is used as the label of the window. This method ensures that the label of the sample is associated with the latest state in the time window. The step size parameter of the sliding window mechanism is set to 1, that is, the window moves one time step each time. This meticulous sliding window method allows subtle changes in time series data to be captured, providing richer data samples, thereby improving the accuracy and reliability of fault detection and diagnosis.

[0057] In another exemplary embodiment, in the above steps 102 and 104, fault detection is performed based on the DAE-C fault detection model, and the fault-free data set is first reconstructed. For the deep embedding clustering network (i.e., the first deep autoencoder), only the data of the fault-free data set is used, and then the trained first deep autoencoder is used to encode the unknown data, in order to distinguish between faulty and fault-free data behaviors, and the faulty data is selected from the middle to add the fault label.

[0058] The above-mentioned DAE-C fault detection model includes: a first deep autoencoder, an encoder, a nonlinear self-attention layer, a DEC clustering layer and a decoder. Figure 6The fault detection process based on the DAE-C fault detection model is shown. After preprocessing, the unknown data set is clustered into two categories, namely fault and non-fault. First, the first deep autoencoder is used to train the non-fault data set to obtain the data pattern under normal working conditions. The left input represents the unknown data set. These data come from multiple sensors of the AHU (air handling unit), including AHU supply air temperature, mixed air temperature, return air temperature, outdoor air flow, supply duct static pressure, supply fan power, return fan power and room temperature. The right part of the input represents the data set with fault labels, which also comes from multiple sensors of the AHU, and the data dimension is consistent with the left input. The features after dimensionality reduction by the first deep autoencoder are input to the nonlinear self-attention layer, which can capture long-range dependencies and important features in the data and improve the expression ability of the model. Then, the features processed by the nonlinear self-attention layer are input to the DEC clustering layer. The DEC clustering layer embeds the data into a new feature space for clustering analysis to distinguish the faulty data. However, the increase in the number of hidden layers will also increase the weight and bias calculation problems of each neuron. This application introduces a back-propagation algorithm to solve this type of computational problem.

[0059] In another exemplary embodiment, the above steps 102 and 103 may be replaced by the following steps 301 to 307 .

[0060] Step 301: Obtain monitoring data (i.e., non-fault data) D in the non-fault data set no_fault ={X1, X2...,XP}. P=30*1440=43,200.

[0061] Step 302: First, analyze the fault-free data set. Input the preprocessed fault-free data set D no_fault , use the first deep autoencoder to extract features from the fault-free data, and extract the fault-free data features Z = f after feature dimension reduction encoder1 (X). The decoder trains the deep autoencoder L by minimizing the reconstruction error reconstruction =||Xf decoder (Z)|| 2 , where L reconstruction is the reconstruction error, X is the fault-free data, Z is the fault-free data feature, and f encoder1 is the first deep autoencoder, f decoder is the decoder, X′=f decoder (Z), X′ is the fault-free reconstructed data.

[0062] Step 303: Extract unknown data (i.e., monitoring data in unknown data sets) features. P = 30 * 1440 = 43,200. Send it to the trained first deep autoencoder f encoder1 Perform feature extraction to extract the features of the monitoring data after dimensionality reduction

[0063] Step 303: Input the reduced dimension monitoring data features A nonlinear self-attention mechanism is used to further process the features through linear transformation Get query Q, key K and value V, W q , W K , W V are the weight matrices for query and key respectively. After that, the nonlinear activation function is calculated, ReLU(x)=max(0,x). Calculate Q′=RELU(Q), K′=RELU(K), V′=RELU(V). After that, the attention weight is calculated, the similarity between query and key is calculated by dot product, and it is scaled and normalized to calculate the attention result Among them, dk is the dimension of query and key vector, and outputs the attention result A.

[0064] Step 305: Input the attention result A. First, calculate the soft assignment probability q ij :q ij =(1+||Ai-u j || 2 ) -1 / ∑ j′ (1+||A i -u j′ || 2 ) -1 Among them, q ij is the soft assignment probability that the i-th monitoring data feature after dimension reduction belongs to cluster center j, A i is the attention result of the monitoring data feature after the i-th dimension reduction, u j is the cluster center j,u j′ is the cluster center j'. Then calculate the target distribution probability p ij . Among them, p ij is the target distribution probability that the i-th monitoring data feature after dimension reduction belongs to cluster center j, q ij is the soft assignment probability that the i-th reduced-dimensional monitoring data feature belongs to cluster center j, q i′j is the soft assignment probability that the i′th reduced-dimensional monitoring data feature belongs to cluster center j, qi ′j′ is the soft assignment probability that the i′th reduced-dimensional monitoring data feature belongs to cluster center j′, q ij′is the soft assignment probability that the i-th reduced-dimensional monitoring data feature belongs to the cluster center j′. Using the soft assignment probability or target distribution probability of each reduced-dimensional monitoring data feature belonging to each cluster center, cluster each reduced-dimensional monitoring data feature to obtain the clustering result C. Using KL divergence: KL j =∑ i p ij ·log(p ij / q ij ) is used as the loss function for optimization; where KL j is the KL divergence of cluster center j.

[0065] Step 306: Reconstruct the data back to the original dimension, that is, use the decoder to reconstruct the reduced-dimensional monitoring data features of the fault type in the clustering result to obtain the fault reconstructed data, D FAULT =f decoder (Z 故障 ) to obtain fault data; where D FAULT Reconstruct data for failure, Z 故障 It is the monitoring data feature after dimensionality reduction of fault type in the clustering result.

[0066] Step 307: End.

[0067] In fault diagnosis, processing time series data is crucial, especially in HVAC systems where the operating status changes over time. Therefore, the model must be able to capture these changes and make accurate judgments. In order to improve the accuracy of fault diagnosis, this application introduces the BIGRU model. The design of the bidirectional GRU layer allows for a more comprehensive capture of forward and backward sequence information during the operation of the HVAC system. Figure 7 shown.

[0068] In order to better process time series data, overlapping time windows are used to capture changes in system states. Each time window contains multiple time points. By processing the data at these time points, the BIGRU model can track and predict the system state more accurately. Within each time window, the core structure of the BIGRU model includes update gates and reset gates, which control how the model integrates current and past information. In order to improve the accuracy of fault diagnosis, it is recognized that different types of faults have different effects on various sensitive features, and 4 implicit fault feature indicators are defined. The definitions and schematic diagrams of these implicit fault features are detailed as follows.

[0069] (1) High-order difference Among them, Δ K x k Reconstruct the kth feature X in the data for the window fault k The higher-order difference of is the binomial coefficient, xk-k′ is the offset from the kth feature to the k′th feature in the window fault reconstruction data. Constructing high-order differences for time series data to achieve stationarity, such as first-order differences or second-order differences, eliminates trends to obtain a stationary representation of the data. This method helps identify abnormal data that deviates from a stationary pattern. Usually, abnormal data is related to faults, and the degree of deviation of the outliers reflects the severity of the fault and helps identify key HVAC faulty components.

[0070] (2) Abnormal Accompanying Among them, A k is the abnormal follow-up of the kth feature in the time window fault reconstruction data, Xk is the kth feature in the time window fault reconstruction data, and ∈k is the threshold corresponding to the kth feature in the time window fault reconstruction data. This refers to whether the change amplitude of a certain feature in the same time window reaches its corresponding threshold. If the threshold is met, the abnormal feature is true; otherwise, it is false. This indicator is used to identify feature combinations with true abnormal features.

[0071] (3) Asynchronous time T k,k′ =t k -t k′ ; Among them, T k,k′ The asynchronous time series of the kth feature and the k'th feature in the time window fault reconstruction data, t k The time when the kth feature in the time window fault reconstruction data becomes abnormal, t k′ Reconstruct the time when the k'th feature in the time window fault data is abnormal. This concept includes calculating the time interval and time companion between synchronous and concurrent feature changes. The time interval refers to the time required for the trend of change between any two features, which can be expressed as a time difference, where a positive value indicates an early change and a negative value indicates a late change.

[0072] (4) Peak time This refers to the time it takes for data to reach an abnormal peak from a stable value. is the abnormal peak time, t start is the abnormal start time. Typically, the peak time is related to the sensitivity of the feature to a specific fault, which helps identify vulnerable or sensitive components in the HVAC system. Considering that there may be multiple peaks, the time difference from the initial change to each peak of each feature is calculated, and the value with the smallest time difference between all peaks is selected as the peak time.

[0073] In another exemplary embodiment, in the above step 104, the above four indicators are combined to summarize a comprehensive judgment rule for each fault type.

[0074] (1) Outdoor air temperature sensor deviation. High-order difference Δ k x iContinuously deviate from normal values ​​and the peak time is shorter.

[0075] (2) Supply air temperature sensor deviation: Similar to the outdoor air temperature sensor deviation, it is determined by high-order difference and peak time.

[0076] (3) The outdoor damper is stuck. k The test shows that the damper opens and remains at a fixed point for a long time, and the damper position cannot be changed within the set time.

[0077] (4) The cooling coil valve is stuck. Abnormality is accompanied by A k Detect cooling coil valve to maintain fixed point for a long time, asynchronous time T i，j The display tube valve signal does not respond to the control signal.

[0078] (5) Cooling coil valve leakage. Asynchronous time T k,k′ The time difference between the actual opening and the control signal increases, and a small but continuous side leakage feature appears.

[0079] Generate fault type label data in each time window according to the rules: Among them, D FAULTTYPE is the fault type label data, D FAULT Reconstruct data for time window failures, Fault type pseudo label.

[0080] In another exemplary embodiment, the above step 105 may be replaced by the following steps 401 to 405 .

[0081] Step 401: A second deep autoencoder with the same architecture as in the DAE-C fault detection model is used to extract low-dimensional features, and a multi-head attention mechanism is introduced in the BIGRU model. The multi-head attention mechanism applies multiple linear transformations to the input, allowing multiple groups of attention scores to be calculated in parallel MultiHead(Q″, K″, V″) = Concat(head1, ..., head n )W0, where Q″, K″ and V″ are the query, key and value in the multi-head attention mechanism, head1, head n They are the attention results of the 1st and nth heads in the multi-head attention mechanism, Concat() represents the connection, and W0 represents the weight matrix in the multi-head attention mechanism.

[0082] Step 402: Reconstruct the time window fault data D FAULT Feed it into the second deep autoencoder Get the fault data features after dimension reduction

[0083] Step 403: Reduce the dimension of the fault data features It is sent to the BIGRU (i.e., bidirectional gated unit), and the update gate first determines how much should be retained in the current hidden state, according to and the final hidden state of the previous time step Calculated by update gate Output in, is the current time step t w The fault data features after dimension reduction, is the previous time step t w The final hidden state of -1, is the current time step t w The output of the update gate, σ is the sigmoid activation function, W z is the weight matrix of the update gate.

[0084] Step 404: Fault data features after dimensionality reduction and the final hidden state of the previous time step Then calculate the degree of combination through the reset gate Output in, is the combination degree of reset gate output, W r is the weight matrix of the reset gate.

[0085] Step 405: Based on the output of the update gate The final hidden state at the previous time step and reset gate output Compute candidate hidden states Output candidate hidden states; where, is the current time step t w The candidate hidden states of is the previous time step t w The final hidden state is -1.

[0086] Step 406: Input Update Gate Candidate hidden states and the final hidden state of the previous time step Compute the final hidden state, combining the candidate hidden state with the hidden state from the previous time step Output the final hidden state; where, is the current time step t w The final hidden state.

[0087] Step 407: Finally, pass through the multi-head attention layer and input the final hidden state Computing multiple independent attention Calculate and concatenate the results MultiHead(Q′, K′, V′)=Concat(head1, ..., head n )W0, and then obtain the final output attention result through linear transformation

[0088] Step 408: Input attention results With the final hidden state Combine the two to output the final feature

[0089] Step 409: The final feature H final Mapped to the category space through the classifier weight W, and generate a predicted distribution is the probability distribution of samples belonging to each category. Through the reverse optimization function Loss = L CON +L REG Continuously optimize the parameters of the DAE-BIGRU fault diagnosis model.

[0090] Among them, L CON is the consistency loss, LREG is the regularization loss, Where N is the number of samples, C is the number of categories, Reconstruct the pseudo-label of the fault type c for the nth time window fault data, The predicted probability that the fault reconstruction data for the nth time window belongs to fault type c.

[0091] Step 410: End.

[0092] In another exemplary embodiment, the present application uses reinforcement learning methods to infer the transmission relationship between faults between components. A directed line between two components represents a transmission relationship. In the fault diagnosis of HVAC systems, data-driven methods, especially methods using reinforcement learning techniques, can accurately identify components related to the fault and find the root cause of the fault without expert knowledge. In addition, these methods can also reveal fault symptom chains.

[0093] The embodiment of the present application focuses on solving the fault chain relationship between the components of the air-conditioning unit. Before building the reinforcement learning framework, the fault data is first extracted and processed using a time window. The reinforcement learning framework is then used to analyze the data and identify the components related to the fault. The reinforcement learning model interacts with the environment and optimizes the decision-making process based on the reward function. By utilizing implicit features such as high-order differences, anomaly tracking, asynchronous timing, and peak time, the rewards obtained are used to continuously adjust the strategy to identify symptom chains.

[0094] First, according to the working principle of HVAC, determine what features are contained in each component, as shown in Table 2. For example, SA_TEMP and SA_TEMPSPT belong to the AHU_COLD_DECK component, and AHU_COLD_DECK belongs to the air handling component category. By monitoring SA_TEMP and SA_TEMPSPT, the operating status of AHU_COLD_DECK and its impact on the entire system can be analyzed. Similarly, by monitoring the SA_TEMP function mapped to the AHU_COLD_DECK component and the CHWC_VLV function mapped to the AHU_Cooling_Coil (air handling component), the impact of cooling system failures on these two components can be inferred, thereby forming a complete symptom chain. Exemplary, the component symptom chain inference schematic diagram of the five types of failures obtained in the embodiment of the present application is shown in FIG. Figure 8 As shown, different colors represent different component functions, and they point to the next component in chronological order, which is marked as the component name.

[0095] Table 2 Component functions and included features

[0096]

[0097]

[0098]

[0099] In another exemplary embodiment, the above steps 106 and 107 may be replaced by the following steps 501 to 507 .

[0100] Step 501: Obtain the fault type label data D after the fault diagnosis of DAE-BIGRU FAULTTYPE .

[0101] Step 502: Design a reinforcement learning framework. Design a reward function The reward function combines four features: high-order differences, abnormal companions, asynchronous time and peak time, α, β, γ, λ, and 1 / T to guide the model to learn to identify components related to failures. k，k′ , 1 / P k are the weight parameters, respectively, representing the abnormal following feature, representing the inverse of the asynchronous timing, representing the inverse of the peak time, being the minimum number of components to be selected, and representing the number of components actually selected.

[0102] Step 503: For data D with the same fault type label (FAULTTYPE_i, i∈[1,5]), FAULTTYPE_i , calculate the four implicit fault features of each feature, and obtain the enhanced data D with fault type labelsrein_i ={D FAULTTYPE_i , {Δ K X k , A k , T k，k′ , P k}}, then the data D rein Send it to reinforcement learning and use the following formula to optimize Q.

[0103]

[0104] Where n is the number of optimization iterations for the enhanced data of each time window, n=0, 1, ..., N-1, N is the total number of optimization iterations for the enhanced data of each time window, S t It refers to the state S in the tth time window t , S t =[{D FAULTTYPE_i , {Δ K X k , A k , T k，k′ , P k}, time window (t)], S t It contains the fault type, feature difference, and time window information. t It means that in state S t The action selected when a t = [{selected feature set}, {selected features corresponding to {Δ K X k , A k , T k，k′ , P k}set}, time window (t)].

[0105] In a time window, {selected feature set} is a feature randomly selected by the algorithm at the beginning. Later, according to the algorithm update, other features that have strong correlation with the features in the existing set are added to the set in sequence. á represents the state S in the t+1th time window t+1 One of the possible actions in the following situation. t is the immediate reward brought by the action at time t, γ is the discount factor parameter, Represents the state S in the t+1th time window t+1 All possible actions The maximum value Q obtained in , α is the learning rate parameter. By continuously optimizing Q, the feature change association sequence is obtained.

[0106] Step 504: According to the order in which the features appear in the feature change association sequence, the components corresponding to each feature are extracted in turn according to the mapping relationship between features and components in Table 2, and a component symptom list is generated, in which each component only appears once in the list. The component symptom list is the order in which the abnormal behaviors of the components occur during the failure process.

[0107] Step 505: Output the symptom chain diagram of the components of each fault type.

[0108] Step 506: End.

[0109] According to the specific embodiments provided in this application, this application has the following technical effects.

[0110] This application effectively detects and diagnoses HVAC system faults by adopting unsupervised learning methods. This research successfully solves the problem of lack of labeled data, making it possible to detect and diagnose HVAC system faults even without clear fault labels.

[0111] A dimensionality reduction method for deep autoencoders with temporal co-representation (i.e., DAE-C fault detection model) is proposed to address the problem of complex features of high-dimensional BAS (Building Automation System) data. The method uses data input in a time window and designs and integrates a nonlinear attention layer to improve the dimensionality reduction effect, thereby generating a clustering discriminator to perform fault detection on new samples. The main features include the ability to efficiently recognize and process large amounts of data, different types of data, and complex representation data, ensuring high accuracy and robustness in fault detection in diverse and complex environments.

[0112] A deep autoencoder-bidirectional gated recurrent unit (DAE-BIGRU fault diagnosis model) method based on a multi-head attention mechanism is proposed. It aims to mine the fault features implicit in overlapping time windows with variable scales, such as high-order differences, change advances, asynchronous times, and peak times, and generate fault prediction labels by combining highly correlated features, and use the DAE-BIGRU method to optimize the labels of fault types.

[0113] In order to achieve fault component diagnosis and fault symptom chain inference, this application proposes a fault implicit feature connection reinforcement learning analysis method to deeply explore the association between faults and feature data and their components. Through the fault symptom chain pattern obtained from the data representation, this strategy has accuracy and can effectively identify the multidimensional data causal chain pattern that causes the fault. This method not only helps fault early warning and equipment maintenance, but also provides more comprehensive fault diagnosis and prevention measures.

[0114] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0115] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for fault detection and diagnosis of HVAC systems based on deep double-layer autoencoder, characterized in that: include: Construct fault-free and unknown datasets for HVAC systems; Training a first deep autoencoder in a DAE-C fault detection model based on the fault-free data set to obtain a trained DAE-C fault detection model; Based on the trained DAE-C fault detection model, fault detection is performed on the monitoring data in the unknown data set to obtain fault reconstruction data; Based on the reconstruction data of each fault, the fault feature linkage analysis strategy is used to generate the time window fault reconstruction data and its fault type pseudo label; According to the fault reconstruction data of each time window and its fault type pseudo-label, the DAE-BIGRU fault diagnosis model is optimized and trained to obtain the trained DAE-BIGRU fault diagnosis model; Based on the trained DAE-BIGRU fault diagnosis model, fault diagnosis of the HVAC system is performed.

2. The HVAC system fault detection and diagnosis method based on deep double-layer autoencoder according to claim 1 is characterized in that: The DAE-C fault detection model includes a first deep autoencoder, an encoder, a nonlinear self-attention layer, a DEC clustering layer and a decoder connected in sequence.

3. The HVAC system fault detection and diagnosis method based on deep double-layer autoencoder according to claim 2 is characterized in that: The method further comprises: training a first deep autoencoder in the DAE-C fault detection model based on the fault-free data set to obtain a trained DAE-C fault detection model; and Using a first deep autoencoder to extract features of the fault-free data in the fault-free data set to obtain features of the fault-free data after dimensionality reduction; Reconstructing the fault-free data features after dimension reduction using the decoder to obtain fault-free reconstructed data; According to the fault-free reconstruction data, the reconstruction error is calculated using the following formula; L reconstruction =||X-X'|| 2 ; Among them, L reconstruction is the reconstruction error, X is the fault-free data, and X′ is the fault-free reconstructed data; Determine whether an iteration end condition is met to obtain a first determination result, wherein the iteration end condition is that the reconstruction error is less than an error threshold or the number of training iterations is not less than an iteration number threshold; If the first judgment result is no, update the parameters of the first deep autoencoder, and return to the step of "using the first deep autoencoder to extract features of the fault-free data in the fault-free data set to obtain features of the fault-free data after dimensionality reduction"; If the first judgment result is yes, the DAE-C fault detection model including the trained first deep autoencoder is output as the trained DAE-C fault detection model; wherein the trained first deep autoencoder is the first deep autoencoder after parameter update in the current iteration process.

4. The HVAC system fault detection and diagnosis method based on deep double-layer autoencoder according to claim 3 is characterized in that: Based on the trained DAE-C fault detection model, fault detection is performed on the monitoring data in the unknown data set to obtain fault reconstruction data, including: Input each monitoring data in the unknown data set into the trained first deep autoencoder to obtain the reduced-dimensional monitoring data features corresponding to each monitoring data; A nonlinear self-attention layer is used to perform nonlinear attention processing on each reduced-dimensional monitoring data feature to obtain the attention results of each reduced-dimensional monitoring data feature; Based on the attention results of each reduced-dimensional monitoring data feature, the DEC clustering layer is used to cluster each reduced-dimensional monitoring data feature to obtain the clustering results; The decoder is used to reconstruct the reduced-dimensional monitoring data features of the fault type in the clustering results to obtain fault reconstructed data.

5. The HVAC system fault detection and diagnosis method based on deep double-layer autoencoder according to claim 4 is characterized in that: Based on the attention results of each reduced-dimensional monitoring data feature, the DEC clustering layer is used to cluster each reduced-dimensional monitoring data feature to obtain the clustering results, including: Initialize cluster centers; Based on the attention results of each reduced-dimensional monitoring data feature, the soft assignment probability of each reduced-dimensional monitoring data feature belonging to each cluster center is calculated using the following formula; q ij =(1+||A i -u j || 2 ) -1 / ∑ j′ (1+||A i -u j′ || 2 ) -1 ; Among them, q ij is the soft assignment probability that the i-th monitoring data feature after dimension reduction belongs to cluster center j, A i is the attention result of the monitoring data feature after the i-th dimension reduction, u j is the cluster center j,u j′ is the cluster center j'; According to the soft assignment probability of each monitoring data feature after dimensionality reduction belonging to each cluster center, the target distribution probability of each monitoring data feature after dimensionality reduction belonging to each cluster center is calculated using the following formula; Among them, p ij is the target distribution probability that the i-th monitoring data feature after dimension reduction belongs to cluster center j, q ij is the soft assignment probability that the i-th reduced-dimensional monitoring data feature belongs to cluster center j, q i′j is the soft assignment probability that the i′th reduced-dimensional monitoring data feature belongs to cluster center j, q i′j′ is the soft assignment probability that the i'th reduced-dimensional monitoring data feature belongs to cluster center j', q ij′ is the soft assignment probability that the i-th monitoring data feature after dimension reduction belongs to cluster center j'; Using the soft assignment probability or target distribution probability of each reduced-dimensional monitoring data feature belonging to each cluster center, clustering each reduced-dimensional monitoring data feature is performed to obtain a clustering result; According to the soft assignment probability and target distribution probability of each monitoring data feature after dimensionality reduction belonging to each cluster center, the KL divergence of each cluster center is calculated using the following formula; KL j =∑ i p ij ·log(p ij / q ij ); Among them, KL j is the KL divergence of cluster center j; Determine whether the KL divergence of each cluster center is less than the divergence threshold, and obtain a second determination result; If the second judgment result is no, the cluster center is updated, and the process returns to the step of "calculating the soft assignment probability that each reduced-dimensional monitoring data feature belongs to each cluster center"; If the second judgment result is yes, the clustering result is output.

6. The HVAC system fault detection and diagnosis method based on deep double-layer autoencoder according to claim 1 is characterized in that: Based on the reconstruction data of each fault, the fault feature linkage analysis strategy is used to generate the time window fault reconstruction data and its fault type pseudo label, including: The reconstruction data of each fault are organized into time series data in chronological order; Dividing the time series data in an overlapping time window manner to obtain time window fault reconstruction data; Calculate the high-order difference, anomaly companion, asynchronous time and peak time of the fault reconstruction data in each time window; Based on the high-order differences, abnormal accompaniment, asynchronous time and peak time of the fault reconstruction data in each time window, a fault feature linkage analysis strategy is adopted to determine the fault type pseudo-label of the fault reconstruction data in each time window.

7. The HVAC system fault detection and diagnosis method based on deep double-layer autoencoder according to claim 1 is characterized in that: The DAE-BIGRU fault diagnosis model includes a second deep autoencoder, a bidirectional gating unit, a multi-head attention mechanism layer, a fully connected layer and an activation function layer connected in sequence.

8. The HVAC system fault detection and diagnosis method based on deep double-layer autoencoder according to claim 7 is characterized in that: According to the fault reconstruction data of each time window and its fault type pseudo label, the DAE-BIGRU fault diagnosis model is optimized and trained to obtain the trained DAE-BIGRU fault diagnosis model, which specifically includes: A second deep autoencoder is used to encode the fault reconstruction data of each time window to obtain the fault data features after dimension reduction corresponding to the fault reconstruction data of each time window; Input each dimension-reduced fault data feature into the bidirectional gating unit to obtain the final hidden state corresponding to the fault reconstruction data in each time window; Use a multi-head attention mechanism layer to calculate the attention results of each final hidden state; Fusing each of the final hidden states and the attention results of the final hidden states to obtain final features corresponding to the fault reconstruction data of each time window; Based on the final features corresponding to the fault reconstruction data of each time window, a fully connected layer and an activation function layer are used to determine the fault type prediction label of the fault reconstruction data of each time window; Based on the fault type prediction labels of the fault reconstruction data in each time window, the fault type pseudo labels of the fault reconstruction data in each time window are continuously optimized through the reverse optimization function to determine the fault type of the fault reconstruction data in each time window.

9. The HVAC system fault detection and diagnosis method based on deep double-layer autoencoder according to claim 8 is characterized in that: The formula of the bidirectional gating unit is expressed as: in, is the current time step t w The fault data features after dimension reduction, is the previous time step t w The final hidden state of -1, is the current time step t w The output of the update gate, σ is the sigmoid activation function, W z is the weight matrix of the update gate, is the combination degree of reset gate output, W r is the weight matrix of the reset gate, W h is the weight matrix of the candidate hidden state, is the current time step t w The candidate hidden states of is the current time step t w The final hidden state.

10. The HVAC system fault detection and diagnosis method based on deep double-layer autoencoder according to claim 1, characterized in that: According to the fault reconstruction data of each time window and its fault type pseudo label, the DAE-BIGRU fault diagnosis model is optimized and trained to obtain the trained DAE-BIGRU fault diagnosis model, which then includes: Based on the fault type of the fault reconstruction data in each time window, the fault symptom chain of the HVAC system is determined by using a reward function optimization decision-making method; Based on the fault symptom chain of the HVAC system, the mapping relationship between features and components is used to determine the order in which each component exhibits abnormal behavior during the fault process, and to construct a component symptom chain diagram for each fault type; The reward function is: Among them, R is the reward function, A k The anomaly following of the kth feature in the time window fault reconstruction data, T k,k′ The asynchronous time series of the kth feature and the k'th feature in the time window fault reconstruction data, P k is the kth characteristic peak time in the time window fault reconstruction data, α, β, γ, λ are all weight parameters, C min is the minimum number of components selected, and |S| is the actual number of components selected.

Images