Intelligent building HVAC system fault detection method based on digital twinning

Abstract

The invention discloses an intelligent building HVAC system fault detection method based on digital twinning, and relates to the technical field of building information management. On the premise that an HVAC data set is obtained, a BoW-based supervised feature extraction and selection method is used, and the HVAC system fault detection method based on digital twinning is obtained. A sensor set only limited to detection of each specified fault can be uniquely selected according to marked normal and fault data, a subsystem special for the specific fault is automatically formed, and manual interference is not needed. Fault tags are correspondingly created in the semantic model, which knowledge is integrated as temporary knowledge into the digital twin ecosystem to associate the most relevant sensor with the corresponding fault tag. The data continuously generated by the sensor carries an associated fault tag, ensuring that the additional processing data is independent and self-contained. And filtering a series of data streams in a JSON object form, finding out a data stream containing a specific fault tag, and alarming the occurrence of a fault according to a goodness-of-fit test.

Description

Technical Field

[0001] The present invention relates to the technical field of building information management, and in particular to a fault detection method for an intelligent building HVAC system based on digital twins. Background Art

[0002] People living in modern society spend a considerable amount of time indoors. The HVAC system in a building is essential to ensure a good indoor environmental quality. Traditional HVAC (heating, ventilation and air conditioning) system fault detection and diagnosis methods usually rely on periodic maintenance and manual inspections, which is not only time-consuming and labor-intensive, but also often difficult to achieve real-time discovery and processing of problems. With the improvement of the degree of building informatization, the development of digital twin technology has provided new solutions for building system management.

[0003] Digital twins, by creating virtual copies of physical assets, can reflect and predict the status and behavior of assets in real time, thereby providing decision support for the operation and maintenance of building systems. Although rich building data carries a huge amount of information, redundant and irrelevant data dimensions can lead to overfitting problems and heavy computational burdens if not used properly. In order to integrate and process massive amounts of real-time data and ensure the validity of data and the accuracy of fault detection, it is necessary to combine the data twin platform with the fault labeling method, filter and generate low-latency data streams, automate the process, and target five functional faults that significantly change the behavior of the HVAC system, including over-infiltration, positive and negative deviations in thermostat measurements in core and peripheral areas; these faults are artificially created and imposed on the facility: over-infiltration is achieved by opening windows to achieve the target infiltration rate, and positive or negative deviations in thermostat measurements are achieved by adjusting the temperature set point in the opposite direction. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a fault detection method for an intelligent building HVAC system based on digital twins, which solves the problems of the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: A fault detection method for an intelligent building HVAC system based on digital twins, comprising:

[0006] Step 1: Fault data collection and preprocessing, using a variety of sensors to monitor and collect the working status data of RTUs and VAVs in the facility;

[0007] Step 2: Feature extraction, fault label definition and real-time monitoring; The BoW-based feature extraction and selection method is used to select the most distinctive sensor group corresponding to each of the different faults; then for each labeled day’s labeled data, it is converted into 277 code words of 4 letters;

[0008] Step 3: The corresponding sensor group and code word selected in step 2 are used to supplement the fault label in the semantic model; and the fault label is added to the Brick model. By combining the sensor group and code word with the semantic model, accurate annotation and fault detection of real-time data streams are achieved;

[0009] Step 4: Fault detection and analysis, filtering the identified sensitive sensory dimensions of a large amount of real-time labeled data; data streams marked with specific fault labels are filtered and exposed through data pipelines to feed into the divergence-driven FDD function; in the implementation process, a WebSocket connection is used to obtain real-time sensor data preprocessed in the digital twin data platform; and the number of times the codewords converted from the dimension-reduced error data stream appear in the past few windows is compared with the probability histogram; the true positive rate of the five fault types is calculated separately through the goodness-of-fit test based on the Kullback-Leibler divergence; the goodness-of-fit test largely detects the target five faults based on the provided labeled data.

[0010] As a further solution of the present invention: in said step 1, the collection and preprocessing of said fault data, using a variety of sensors to monitor and collect the working status data of the RTU and VAV in the facility, specifically includes:

[0011] S1: First, the labeled data of RTU and VAV under normal and fault conditions are sampled to five minutes to solve the problem of missing sensor data values;

[0012] The labeled data includes data collected by various collection sensors;

[0013] S2: After the data preprocessing process, the labeled data is sliced ​​in an overlapping manner using a sliding window of size 12;

[0014] S3: Labeled data of 4 normal days and 1 fault day are used to identify the sensitive sensory dimensions of each fault type.

[0015] As a further solution of the present invention: the slicing method includes:

[0016] Starting from the data sequence of the labeled data, take the first 12 data points as a window; then move the window forward a certain step size and take the next 12 data points as the next window.

[0017] As a further solution of the present invention: in the step 2, the specific content of the feature extraction and selection method based on BoW is:

[0018] The normal and fault data of the labeled day are used to identify the sensitive sensory dimensions of the fault, thereby supporting subsystem decomposition to detect the target HVAC fault. SAX in BoW is used to extract features from HVAC sensor data. The feature extraction method based on BoW uses SAX technology and feature selection method of χ2 test to extract useful features from the sensor data of the HVAC system; identify and diagnose faults in the HVAC system.

[0019] As a further solution of the present invention: the SAX in the BoW is used to extract features from the HVAC sensor data, and the specific contents of the feature extraction method based on the BoW using the SAX technology and the feature selection method of the χ2 test to extract useful features from the sensor data of the HVAC system include;

[0020] For the normalized time series data x from sensor i i =[x 1i ,x 2i ,x 3i ] T , a sliding window of length w (w≤t) is used to segment the i-th sensor data; assuming the k-th observation result is The observations are further divided into segments of size p, where p represents the codeword length;

[0021] The codeword histogram is used as a criterion for feature selection, and the χ2 test is used to determine the consistency of the codeword histogram and identify the codeword most relevant to the fault; the χ2 score of the corresponding codeword is calculated as follows:

[0022]

[0023] T and P represent the counts of a specific codeword and all codewords found in error data, M and Z represent the counts of a specific codeword and all codewords found in normal and error data;

[0024] The χ2 test is performed to find the most discriminative codewords for distinguishing normal and different fault conditions; the sensory dimensions that generate corresponding codewords and have a χ2 score greater than a threshold will be selected for fault detection.

[0025] As a further solution of the present invention: in the step 3, the selected corresponding sensor group and code word are used to supplement the fault label in the semantic model; and the fault label is added to the Brick model, and the content of accurate annotation and fault detection of real-time data stream is realized by combining the sensor group and code word with the semantic model;

[0026] Fault labels are supplemented in the corresponding semantic models to retain this temporary knowledge in the digital twin ecosystem; the Brick model of the HVAC system in the facility defined using the Brick ontology defines the physical and logical entities in the facility, and the relationships are modeled to capture the connections between these entities; for the five fault types, five labels are added to the Brick model, namely fault_a, fault_b, fault_c, fault_d, and fault_e, and the sensors that provide information for each fault are marked.

[0027] As a further solution of the present invention: the selected corresponding sensor group and code word are used to supplement the fault label in the semantic model; and the fault label is added to the Brick model, and the content of realizing accurate annotation and fault detection of real-time data stream by combining the sensor group and code word with the semantic model also includes;

[0028] Using fault labels, the sensitive sensory dimensions of each fault are uniquely labeled; in addition, the digital twin data platform also annotates minute-level data streams with metadata including defined fault labels;

[0029] Convert real-time data streams into JSON format to ensure real-time data integration; Sliced ​​Brick models can be converted into crates, where the hasTag, hasPoint connections in the Brick model are highlighted; The connection contains all the metadata required to annotate the real-time data stream;

[0030] Convert the Brick model to a crate data model and annotate the metadata of the corresponding device, location, and fault label on the real-time data stream; finally, filter the real-time data stored as a sequence of JSON objects, and then expose and publish the data containing the target fault label to the FDD function, thereby driving the FDD function using Kullback-Leibler divergence;

[0031] When obtaining symbol sequences that are more easily distinguishable between normal conditions and fault conditions; by converting the time series into symbolic representation, the low-dimensional symbol sequence can be effectively coupled with the goodness-of-fit test based on Kullback-Leibler divergence to reveal the occurrence of specific faults; Kullback–Leibler divergence is defined as follows:

[0032]

[0033] and Q represent the probability histograms of these codewords generated by normal data and observed data, respectively; a goodness-of-fit test based on Kullback-Leibler divergence is used to classify the sliding time series as erroneous or normal by tracking the time evolution distribution of the generated codewords.

[0034] The present invention provides a fault detection method for an intelligent building HVAC system based on digital twins. Compared with the prior art, it has the following beneficial effects:

[0035] Under the premise of acquiring HVAC data sets, the present invention, with the help of a supervised feature extraction and selection method based on BoW, can automatically form a subsystem dedicated to a specific fault without manual intervention by only selecting a set of sensors that detect each specified fault based on the labeled normal and fault data; create fault labels in the semantic model accordingly, and integrate this knowledge into the digital twin ecosystem as temporary knowledge to associate the most relevant sensors with the corresponding fault labels; the data continuously generated by the sensor carries the relevant fault labels, ensuring that the attached processed data is independent and self-contained; filter a series of data streams in the form of JSON objects, find out the data streams containing specific fault labels, and alarm the occurrence of faults based on the goodness of fit test;

[0036] The present invention introduces a SAX-based feature extraction and selection method as a typical symbolic artificial intelligence technology to improve computing efficiency; "fault tags" are defined accordingly in the Brick model to mark these related sensor entities; a digital twin data platform is adopted to integrate the defined knowledge tags with real-time data; for the future of digital twins, the proposed framework can overcome the problem of "big" data (large capacity, multiple varieties, but not fast) and filter a large amount of data from the digital twin ecosystem; due to the metadata tagging and annotation method, the developed method can be replicated and allows full use of ontological knowledge and temporary knowledge in artificial intelligence technology to support smart building functions; the expressiveness and richness of semantics are used to develop complex functions, thereby better expressing a comprehensive understanding of building systems and promoting real-time management of physical assets; the data integration method is based on the digital twin data platform, giving priority to the real-time nature of data, and treating pseudo-static information as metadata attached to real-time data streams; it makes this method flexible and very effective for the development of dynamic smart building functions (such as asset management applications). BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The present invention will be further described below in conjunction with the accompanying drawings.

[0038] Figure 1 This is a flowchart of the steps of a digital twin-based intelligent building HVAC system fault detection method of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] Embodiment 1

[0041] The present invention targets five functional faults that significantly change the behavior of the HVAC system, including over-infiltration, positive and negative deviations in thermostat measurements in the core and perimeter zones; these faults are artificially created and imposed on the facility: over-infiltration is achieved by opening windows to achieve the target infiltration rate, positive or negative deviations in thermostat measurements are achieved by adjusting the temperature set point in the opposite direction;

[0042] See also Figure 1 ,The present invention provides a fault detection method for an intelligent building HVAC system based on digital twin, comprising;

[0043] Step 1: Fault data collection and preprocessing, using a variety of sensors to monitor and collect the working status data of the RTU (remote terminal unit) and VAV (variable air volume device) in the facility;

[0044] Specifically, the multiple sensors include temperature, humidity, pressure, airflow, and energy consumption (electricity and gas) sensors. In this embodiment, data collection is performed at a sampling interval of one minute;

[0045] The collection and preprocessing of the fault data, using a variety of sensors to monitor and collect the working status data of the RTU and VAV in the facility, specifically includes:

[0046] S1: First, the tag data of RTU and VAV under normal and fault conditions are sampled to five minutes to solve the problem of missing sensor data values;

[0047] The labeled data includes data collected by various collection sensors, such as HVAC system sensor data, temperature, humidity sensors, etc.;

[0048] It should be noted that downsampling to five minutes can solve the problem of missing sensor data values. It smoothes the data by reducing the number of data points, reduces the data volume and reduces the processing burden, and estimates missing values ​​through interpolation, thereby reducing the impact of noise and simplifying data analysis. The specific process includes data collection, time alignment, downsampling, missing data processing and verification;

[0049] S2: After the data preprocessing process, the labeled data is sliced ​​in an overlapping manner using a sliding window of size 12;

[0050] It should be noted that sliding window refers to a method of continuously taking fixed-length data segments for analysis at a certain step size on a certain time series data; a window size of 12 means that each analyzed data segment contains 12 consecutive data points; the overlapping method means that when the window moves on the data sequence, a part of the data between each window is overlapped;

[0051] The slicing method includes: starting from the data sequence of the labeled data, taking the first 12 data points as a window; then moving the window forward by a certain step size (for example, 6 data points each time), and taking the next 12 data points as the next window;

[0052] S3: Labeled data of 4 normal days and 1 fault day are used to identify the sensitive sensory dimensions of each fault type;

[0053] The sensitive sensory dimension can be understood as the sensors related to fault detection and diagnosis in the HVAC system, which are used to identify and distinguish different fault types (such as temperature, humidity, etc.);

[0054] It should be noted that step S3 is a part of the entire data processing and analysis process, which illustrates the use of labeled data from four normal days and one faulty day and the use of these data to identify sensitive sensory dimensions; different sensitive sensory dimensions mean different sensor data types;

[0055] Step 2: Feature extraction, fault label definition and real-time monitoring; BoW (bag of words) based feature extraction and selection method is used to select the most distinctive sensor group corresponding to each of the different faults; then for each labeled day’s labeled data, it is converted into 277 code words of 4 letters;

[0056] It should be noted that the process of labeled data conversion is part of the BoW-based feature extraction method. In BoW-based feature extraction, data is usually converted into discrete symbolic representations, which can be used to train machine learning models;

[0057] Specifically, “most discriminating” refers to the features that are most effective in distinguishing between the different categories (normal operation and five different fault states);

[0058] Specifically, the marking data of each marking day is converted into 277 unique code words consisting of 4 letters, which are based on a dictionary containing 4 letters; the two 4s here represent the size of the dictionary and the length of the code word, which are both 4; the code word is a fixed-length and unique symbol sequence generated during the encoding process, which is used for information representation and transmission;

[0059] The specific contents of the feature extraction and selection method based on BoW (bag of words) are as follows:

[0060] The normal and fault data of the labeled day are used to identify the sensitive sensory dimensions of the fault, thereby supporting the subsystem decomposition to detect the target HVAC fault. SAX (Symbolic Aggregate Approximation) in BoW is used to extract features from the HVAC sensor data. The feature extraction method based on BoW uses the SAX technique and the feature selection method of χ2 (Chi-square) test to extract useful features from the sensor data of the HVAC system; these features will be used to train the fault detection model to identify and diagnose faults in the HVAC system.

[0061] SAX (Symbolic Aggregate Approximation) in the BoW is used to extract features from HVAC sensor data. The BoW-based feature extraction method uses SAX technology and a feature selection method of χ2 (chi-square) test to extract useful features from the sensor data of the HVAC system. Specifically, the content includes:

[0062] Feature selection method based on SAX and χ2 (chi-square) test, which is used to extract features from normalized time series data for fault detection:

[0063] For the normalized time series data x from sensor i i =[x 1i ,x 2i ,x 3i ] T , a sliding window of length w (w≤t) is used to segment the i-th sensor data; assuming the k-th observation result is The observations are further divided into segments of size p, where p represents the codeword length;

[0064] The codeword histogram is used as a criterion for feature selection, and the χ2 test is used to determine the consistency of the codeword histogram and identify the codeword most relevant to the fault; the χ2 score of the corresponding codeword is calculated as follows:

[0065]

[0066] T and P represent the counts of a specific codeword and all codewords found in error data, M and Z represent the counts of a specific codeword and all codewords found in normal and error data;

[0067] The χ2 test is performed to find the most discriminative codewords that distinguish normal and different fault conditions; sensory dimensions that generate corresponding codewords and have a χ2 score greater than the threshold will be selected for fault detection;

[0068] Step 3: The corresponding sensor group and code word selected in step 2 are used to supplement the fault label in the semantic model; and the fault label is added to the Brick model. By combining the sensor group and code word with the semantic model, accurate annotation and fault detection of real-time data streams are achieved;

[0069] It should be noted that the Brick model defines the physical and logical entities of the HVAC system in the facility and the relationships between them; for each fault type, the corresponding label is added to the Brick model, and the sensor that provides the fault information is marked; the real-time data stream is annotated, including sensor readings, location information, and fault labels, which are converted into JSON format to support real-time data integration; the sliced ​​Brick model is converted into crates, which contains all the metadata required for real-time data stream annotation;

[0070] The selected corresponding sensor groups and code words are used to supplement the fault labels in the semantic model; and the fault labels are added to the Brick model, and the accurate annotation and fault detection of the real-time data stream are achieved by combining the sensor groups and code words with the semantic model;

[0071] Supplement fault labels in the corresponding semantic models to preserve this temporary knowledge in the digital twin ecosystem; use the Brick model of the HVAC system in the facility defined by the Brick ontology to define the physical and logical entities in the facility (such as areas, RTUs, VAVs, their components, and deployed sensors), and model relationships (such as hasPart, feeds, hasPoint) to capture the connections between these entities; for the five fault types, five labels (fault_a, fault_b, fault_c, fault_d, fault_e) are added to the Brick model, and the sensor that provides information for each fault is marked (i.e. hasTag, failure_x);

[0072] The selected corresponding sensor groups and code words are used to supplement the fault labels in the semantic model; and the fault labels are added to the Brick model, and the accurate annotation and fault detection of the real-time data stream are realized by combining the sensor groups and code words with the semantic model;

[0073] Using fault labels, the sensitive sensory dimension of each fault is uniquely labeled. In addition, the digital twin data platform also annotates minute-level data streams with metadata including defined fault labels. Taking fault_d as an example, the real-time readings of the five tagged sensors are annotated with the relevant equipment, location, and especially the fault label (fault_d) to create independent messages that feed into the FDD function. Specifically, for the sensor reading with the point identifier ZoneT_205, the location of the sensor (Zone_205) and the associated fault label (fault_d) are attached and annotated to the JSON object.

[0074] Convert real-time data streams into JSON format to ensure real-time data integration; sliced ​​Brick models can be converted into crates, where the hasTag, hasPoint connections in the Brick model are highlighted; these connections contain all the metadata required to annotate real-time data streams;

[0075] Furthermore, the Brick model is converted into a crate data model, and metadata of the corresponding device, location, and fault labels are annotated on the real-time data stream; finally, the real-time data stored as a sequence of JSON objects can be filtered, and then the data containing the target fault labels can be exposed and published to the FDD function, thereby driving the FDD function using Kullback-Leibler divergence;

[0076] Furthermore, considering the need to obtain symbol sequences that are more easily distinguishable between normal conditions and fault conditions; by converting the time series into symbolic representation, the low-dimensional symbol sequence can be effectively coupled with the goodness-of-fit test based on Kullback-Leibler divergence to reveal the occurrence of specific faults; Kullback–Leibler divergence is defined as follows:

[0077]

[0078] and Q represent the probability histograms of these codewords generated by normal data and observed data, respectively; a goodness-of-fit test based on Kullback-Leibler divergence is used to classify the sliding time series as erroneous or normal by tracking the time evolution distribution of the generated codewords;

[0079] Step 4: Fault detection and analysis, filtering the identified sensory dimensions of a large amount of real-time data; the data stream annotated with specific fault labels is filtered and exposed through the data pipeline to feed into the divergence-driven FDD function; in the implementation process, a WebSocket connection is used to obtain real-time sensor data preprocessed in the digital twin data platform; in order to demonstrate the application of the one-sided goodness-of-fit test in revealing error scenarios, the number of times the codeword converted from the dimension-reduced error data stream appears in the past few windows (i.e. ) is compared with the probability histogram Q; the goodness of fit test based on Kullback-Leibler divergence is used to calculate the true positive rates of the five fault types respectively; the goodness of fit test can detect the target five faults to a large extent based on the provided data.

[0080] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0081] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A fault detection method for HVAC system in intelligent buildings based on digital twins, characterized in that: include: Step 1: Fault data collection and preprocessing, using a variety of sensors to monitor and collect the working status data of RTUs and VAVs in the facility; Step 2: Feature extraction, fault label definition and real-time monitoring; The BoW-based feature extraction and selection method is used to select the most distinctive sensor group corresponding to each of the different faults; then for each labeled day’s labeled data, it is converted into 277 code words of 4 letters; Step 3: The corresponding sensor group and code word selected in step 2 are used to supplement the fault label in the semantic model; and the fault label is added to the Brick model. By combining the sensor group and code word with the semantic model, accurate annotation and fault detection of real-time data streams are achieved; Step 4: Fault detection and analysis, filtering the identified sensitive sensory dimensions of a large amount of real-time labeled data; the data stream marked with specific fault labels is filtered and exposed through the data pipeline to feed into the divergence-driven FDD function; In the implementation process, WebSocket connection is used to obtain real-time sensor data preprocessed in the digital twin data platform; The scores of times the codewords converted from the dimensionality-reduced error data stream appeared in the past few windows are compared with the probability histogram. The true positive rates of the five fault types are calculated respectively through a goodness-of-fit test based on the Kullback-Leibler divergence. The goodness-of-fit test largely detects the five target faults based on the provided labeled data.

2. According to claim 1, a digital twin-based intelligent building HVAC system fault detection method is characterized in that: In the step 1, the collection and preprocessing of the fault data, using a variety of sensors to monitor and collect the working status data of the RTU and VAV in the facility, specifically includes: S1: First, the tag data of RTU and VAV under normal and fault conditions are sampled to five minutes to solve the problem of missing sensor data values; The labeled data includes data collected by various collection sensors; S2: After the data preprocessing process, the labeled data is sliced ​​in an overlapping manner using a sliding window of size 12; S3: Labeled data of 4 normal days and 1 fault day are used to identify the sensitive sensory dimensions of each fault type.

3. The method for fault detection of HVAC system in intelligent building based on digital twin according to claim 2, characterized in that: The slicing methods include: Starting from the data sequence of the labeled data, take the first 12 data points as a window; then move the window forward a certain step size and take the next 12 data points as the next window.

4. The method for fault detection of HVAC system in intelligent building based on digital twin according to claim 1, characterized in that: In the step 2, the specific content of the feature extraction and selection method based on BoW is: The normal and fault data of the labeled day are used to identify the sensitive sensory dimensions of the fault, thereby supporting the subsystem decomposition to detect the target HVAC fault. SAX in BoW is used to extract features from the HVAC sensor data. The feature extraction method based on BoW uses the SAX technique and the feature selection method of the χ2 test to extract useful features from the sensor data of the HVAC system; Identify and diagnose faults in HVAC systems.

5. The method for fault detection of HVAC system in intelligent building based on digital twin according to claim 4, characterized in that: The SAX in the BoW is used to extract features from the HVAC sensor data. The feature extraction method based on the BoW uses the SAX technique and the feature selection method of the χ2 test to extract useful features from the sensor data of the HVAC system. Specifically, the content includes: For the normalized time series data x from sensor i i =[x 1i ,x 2i ,x 3i ] T , a sliding window of length w (w ≤ t) is used to segment the i-th sensor data; assuming the k-th observation result is [x (tk-w+1)i ,x (tk-w+2)i ,...,x tki ] T , the observations are further divided into segments of size p, where p represents the codeword length; The codeword histogram is used as a criterion for feature selection, and the χ2 test is used to determine the consistency of the codeword histogram and identify the codeword most relevant to the fault; the χ2 score of the corresponding codeword is calculated as follows: T and P represent the counts of a specific codeword and all codewords found in error data, and M and Z represent the counts of a specific codeword and all codewords found in normal and error data; The χ2 test is performed to find the most discriminative codewords for distinguishing normal and different fault conditions; sensory dimensions that generate corresponding codewords and have a χ2 score greater than a threshold will be selected for fault detection.

6. The method for fault detection of HVAC system in intelligent building based on digital twin according to claim 1, characterized in that: In the step 3, the selected corresponding sensor group and code word are used to supplement the fault label in the semantic model; and the fault label is added to the Brick model, and the content of accurate annotation and fault detection of real-time data stream is realized by combining the sensor group and code word with the semantic model; Fault labels are supplemented in the corresponding semantic models to retain this temporary knowledge in the digital twin ecosystem; the Brick model of the HVAC system in the facility defined using the Brick ontology defines the physical and logical entities in the facility, and the relationships are modeled to capture the connections between these entities; for the five fault types, five labels are added to the Brick model, namely fault_a, fault_b, fault_c, fault_d, and fault_e, and the sensors that provide information for each fault are marked.

7. The method for fault detection of HVAC system in intelligent building based on digital twin according to claim 6, characterized in that: The selected corresponding sensor groups and code words are used to supplement the fault labels in the semantic model; and the fault labels are added to the Brick model, and the accurate annotation and fault detection of the real-time data stream are realized by combining the sensor groups and code words with the semantic model; Using fault labels, the sensitive sensory dimension of each fault is uniquely labeled; In addition, the digital twin data platform annotates minute-level data streams with metadata including defined fault labels; Convert real-time data streams into JSON format to ensure real-time data integration; Sliced ​​Brick models can be converted into crates, where the hasTag, hasPoint connections in the Brick model are highlighted; The connection contains all the metadata required to annotate the real-time data stream; Convert the Brick model to a crate data model and annotate the metadata of the corresponding device, location, and fault label on the real-time data stream; finally, filter the real-time data stored as a sequence of JSON objects, and then expose and publish the data containing the target fault label to the FDD function, thereby driving the FDD function using Kullback-Leibler divergence; When obtaining symbol sequences that are more easily distinguishable between normal conditions and fault conditions; by converting the time series into symbolic representation, the low-dimensional symbol sequence can be effectively coupled with the goodness-of-fit test based on Kullback-Leibler divergence to reveal the occurrence of specific faults; Kullback–Leibler divergence is defined as follows: and Q represent the probability histograms of these codewords generated by normal data and observed data, respectively; a goodness-of-fit test based on Kullback-Leibler divergence is used to classify the sliding time series as erroneous or normal by tracking the time evolution distribution of the generated codewords.

Images