Building equipment fault positioning method based on sensor network and layer model

Abstract

The invention relates to the technical field of building automation and intelligent operation and maintenance, in particular to a building equipment fault positioning method based on a sensor network and a layer model, and the method comprises the steps: data collection and preprocessing, dual-time scale concept drift detection, dynamic model management and adaptive correction, and fault positioning and output. According to the method, system changes with sudden change and slow change can be captured at the same time through the double-time-scale design, the dynamic model pool and the historical concept reproduction recognition mechanism enable the system to memorize and reuse the historical operation mode, false alarms caused by periodic changes such as season switching are avoided, and the reliability of the system is improved. The problems that a traditional static or simple online updating model is poor in adaptability in a dynamic environment and the accuracy rate is reduced are fundamentally solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of building automation and intelligent operation and maintenance, and in particular to a building equipment fault positioning method based on a sensor network and a layer model. BACKGROUND

[0002] Modern large-scale buildings, such as commercial complexes, office buildings, and hospitals, rely heavily on complex mechanical and electrical equipment systems, especially heating, ventilation, and air conditioning systems, for the comfort, safety, and energy efficiency of their internal environment. In order to achieve fine management and fault warning of these systems, building automation systems and large-scale sensor networks are widely deployed, continuously generating massive amounts of multi-source heterogeneous time series data, such as temperature, humidity, pressure, flow, current, vibration, etc. Using this data to build machine learning or deep learning-based fault diagnosis models has become a key technical path to improve operational efficiency, reduce energy consumption, and prevent major equipment damage.

[0003] However, the operating environment of building equipment is not constant, and its data feature distribution changes dynamically with various factors, a phenomenon known as concept drift in the field of data mining. For example, seasonal changes such as switching between cooling in summer and heating in winter, natural aging and performance degradation of equipment over time, changes in personnel activity patterns within the building, and adjustments to management strategies can all cause changes in the data patterns corresponding to normal operating states of equipment. An air conditioning fault diagnosis model trained in summer may produce a large number of false positives in winter due to its inability to recognize new normal operating patterns. The existence of such concept drift is a major technical bottleneck that leads to poor adaptability and significant decline in diagnostic accuracy over time of existing static fault diagnosis models in long-term practical applications.

[0004] To address this challenge, the academic and industrial communities have made some explorations. However, existing technologies, whether simple online updating mechanisms or purely theoretical algorithm research, have not effectively addressed the main shortcomings of poor adaptability and declining accuracy of building equipment fault diagnosis models due to concept drift. This has resulted in insufficient reliability of existing diagnostic systems in dynamic real-world building environments over long-term operation, failing to meet the requirements of intelligent operation and maintenance for high precision and high stability. SUMMARY

[0005] The application aims to provide a building equipment fault positioning method based on a sensor network and a layer model, comprising: data acquisition and preprocessing: continuously acquiring multi-source heterogeneous time series data from a sensor network covering building equipment, and integrating and standardizing the time series data based on a preset building equipment layer model; double-time scale concept drift detection: a short-term sliding window for detecting sudden drift of data flow and a long-term baseline window for detecting slow or recurring drift of data flow are set in parallel, and the data feature distribution in the short-term sliding window and the long-term baseline window is continuously monitored to identify concept drift events occurring in the building equipment operation state and determine the type thereof; dynamic model management and adaptive correction: a dynamic model pool containing a plurality of historical diagnostic models corresponding to historical concepts is constructed and maintained; when the concept drift event is detected, the following adaptive correction logic is executed: if the concept drift event is identified as a historical concept recurrence, the corresponding historical diagnostic model is retrieved from the dynamic model pool and activated as a current diagnostic model; if the concept drift event is identified as a new concept, an incremental learning mechanism is triggered to generate a new diagnostic model online based on the existing model knowledge; fault positioning and output: the historical diagnostic model or the new diagnostic model is used to analyze the real-time collected time series data to identify abnormal patterns and perform fault positioning, and finally a diagnostic report containing fault information is generated and output.

[0006] Optionally, the short-term sliding window is used to capture rapid data distribution changes caused by equipment sudden failure or sharp change in operation mode, and the long-term baseline window is used to capture slow data distribution changes caused by seasonal change, equipment aging or gradual change in operation load.

[0007] Optionally, in the short-term sliding window, a statistical process control method is used to monitor the instantaneous change of key performance indicators to detect sudden drift; in the long-term baseline window, a data distribution difference test method is used to detect slow drift by comparing the overall data distribution difference between the current window and the historical window.

[0008] Optionally, each historical diagnostic model in the dynamic model pool is associated with a metadata tag describing the historical concept corresponding thereto, and the metadata tag includes but is not limited to applicable season, working condition mode, equipment configuration or timestamp information.

[0009] Optionally, the concept drift event is identified as a historical concept recurrence, comprising: extracting a feature vector of the current time series data, and calculating the similarity between the feature vector and the historical concept features described by the metadata tags of the historical diagnostic models in the dynamic model pool; when the similarity exceeds a preset threshold, it is determined that the historical concept is recurring.

[0010] Optionally, the incremental learning mechanism adopts knowledge distillation or parameter fine-tuning techniques, and trains the diagnosis model using the time series data corresponding to the new concept by constructing a composite loss function containing a prediction loss and a knowledge preservation loss, thereby achieving online updating without complete retraining to avoid catastrophic forgetting of historical knowledge.

[0011] Optionally, the integration and standardization of the time series data based on the preset building equipment layer model includes: mapping the collected sensor data to the corresponding nodes in the equipment layer, subsystem layer and system layer, and performing time alignment, missing value filling and normalization operation on the time series data of different sources to form feature input of unified format.

[0012] Optionally, the fault locating and outputting step further includes: matching and reasoning the abnormal pattern or abnormal score output by the historical diagnosis model or the new diagnosis model with the pre-constructed fault knowledge graph to determine the root cause of the fault.

[0013] Optionally, the diagnosis report includes the device location where the fault occurs, the most likely root cause, the evaluated fault severity level and the recommended maintenance operation suggestion.

[0014] The second object of the present application is to provide a building equipment fault locating system based on sensor network and layer model, comprising: a data acquisition module configured to continuously acquire multi-source heterogeneous time series data from a sensor network covering building equipment, and integrate and standardize the time series data based on a preset building equipment layer model; a concept drift detection module configured to set a short-term sliding window for detecting sudden drift of data flow and a long-term baseline window for detecting slow drift or recurring drift of data flow in parallel, and continuously monitor the distribution of data features in the short-term sliding window and the long-term baseline window to identify concept drift events occurring in the running state of building equipment and determine their types; a model management module configured to build and maintain a dynamic model pool containing a plurality of historical diagnosis models corresponding to historical concepts; when the concept drift event is detected, the following adaptive correction logic is executed: if the concept drift event is identified as a recurring historical concept, the corresponding historical diagnosis model is retrieved from the dynamic model pool and activated as the current diagnosis model; if the concept drift event is identified as a new concept, an incremental learning mechanism is triggered to generate a new diagnosis model online based on the existing model knowledge; a fault locating module configured to analyze the real-time collected time series data using the historical diagnosis model or the new diagnosis model to identify abnormal patterns and locate faults, and finally generate and output a diagnosis report containing fault information.

[0015] The application can capture system changes of both mutation and slow change through double time scale design, and the dynamic model pool and historical concept reproduction identification mechanism enable the system to remember and reuse historical operation modes, avoid false alarms caused by periodic changes such as seasonal switching, and fundamentally solve the problems of poor adaptability and decreased accuracy of traditional static or simply online updated models in dynamic environment. By combining the abnormal information of the model output with the fault knowledge graph for reasoning, the application can further go to the level of "what is the root cause of the fault" instead of only stopping at the level of "whether a fault occurs", and give specific maintenance suggestions, which provides more valuable decision support for operation and maintenance personnel, helps to realize the change from passive maintenance to predictive maintenance, and thus effectively reduces equipment downtime, prolongs equipment life, and reduces total operation and maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The application is a building equipment fault positioning system based on sensor network and layer model.

[0017] Figure 2 The application is a building equipment fault positioning method based on sensor network and layer model. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the application clearer, the technical scheme of the application will be described in detail below with specific embodiments. It should be noted that the specific embodiments herein are only used to explain the application, and do not limit the scope of the application. Those skilled in the art can make various modifications and improvements to the application without departing from the spirit and essence of the application, and these modifications and improvements should also fall within the protection scope of the application.

[0019] As shown in Figure 1 The application discloses a building equipment fault positioning system based on sensor network and layer model, which comprises a data acquisition module, a concept drift detection module, a model management module and a fault positioning module.

[0020] Specifically, the data acquisition module is responsible for interfacing with the sensor network and building automation system on site, continuously collecting original time series data from hundreds of physical or virtual measurement points, and using a preset building equipment layer model to structure process these multi-source heterogeneous time series data to form feature data streams in a unified format for use by subsequent modules.

[0021] The concept drift detection module monitors the feature data stream output by the data acquisition module in real time, and internally runs two detection mechanisms in parallel: one based on a short-term sliding window for quickly capturing sudden changes, and the other based on a long-term baseline window for identifying slow cumulative changes. The output of the concept drift detection module is a judgment signal about whether a concept drift has occurred and the type of drift, such as: no drift, sudden drift, slow drift, historical concept recurrence, etc.

[0022] The model management module receives signals from the concept drift detection module and makes intelligent decisions accordingly. Internally, it includes a dynamic model pool and an incremental learning unit. The dynamic model pool is a special database that stores multiple historical diagnostic models that have been verified for good performance. Each diagnostic model is accompanied by detailed metadata tags describing its applicable working conditions, like a model archive. The incremental learning unit is a functional unit responsible for online updating of diagnostic models. When new knowledge needs to be learned, the incremental learning unit is activated to iterate on existing diagnostic models or generate new diagnostic models using incremental learning algorithms.

[0023] The fault location module obtains the most suitable diagnostic model from the model management module, analyzes real-time time series data using the diagnostic model, identifies abnormal patterns, and once an anomaly is found, calls the pre-built fault knowledge graph for root cause reasoning, and finally generates a highly actionable diagnostic report, which is presented to management personnel through an operation and maintenance platform such as a computer or mobile application.

[0024] As shown in Figure 2 The building equipment fault location method based on a sensor network and a layer model disclosed by the embodiments of the present application includes the following steps.

[0025] S01: Data acquisition and preprocessing: continuously acquire multi-source heterogeneous time series data from the sensor network covering the building equipment, and integrate and standardize the time series data based on the preset building equipment layer model.

[0026] Specifically, the data acquisition module obtains raw data from the sensor network and the building automation system. The sensor network includes temperature, pressure, flow sensors, etc., and the raw data of the building automation system includes device start-stop status, set value parameters, etc. These data are mapped into the building equipment layer model, and after time alignment, missing value filling, normalization, etc. preprocessing steps, standardized multi-dimensional time series feature vectors are formed.

[0027] S02: Concept Drift Detection: A short-term sliding window is set in parallel to detect abrupt drift of data stream, and a long-term baseline window is set to detect slow drift or recurring drift of data stream. By continuously monitoring the distribution of data features in the short-term sliding window and the long-term baseline window, concept drift events occurring in the building equipment operation state are identified and their types are determined.

[0028] It can be understood that the standardized data stream is simultaneously sent to the short-term sliding window and the long-term baseline window of the concept drift detection module.

[0029] S03: Dynamic Model Management and Adaptive Correction: A dynamic model pool containing multiple historical diagnostic models corresponding to historical concepts is constructed and maintained. When a concept drift event is detected, the following adaptive correction logic is executed: if the concept drift event is identified as a recurring historical concept, the corresponding historical diagnostic model is retrieved from the dynamic model pool and activated as the current diagnostic model; if the concept drift event is identified as a new concept, an incremental learning mechanism is triggered to generate a new diagnostic model online based on the existing model knowledge.

[0030] It can be understood that the data acquisition module continuously calculates and compares the distribution of data features in the window. If no significant change is detected, the process returns to step S01, and the system continues to use the current diagnostic model for monitoring. If a concept drift is detected, the data acquisition module sends a drift type (abrupt, slow, recurring, etc.) signal to the model management module. The model management module makes a decision based on the signal: if it is a recurring historical concept, it searches for the most similar historical diagnostic model based on the current data features in the dynamic model pool and activates it as the current diagnostic model; if it is a new concept, it triggers the incremental learning unit to update the current diagnostic model or generate a new diagnostic model online using new data. The new diagnostic model, after being verified, will be stored in the dynamic model pool and labeled with new metadata.

[0031] S04: Fault Location and Output: The real-time collected time series data is analyzed using the historical diagnostic model or the new diagnostic model to identify abnormal patterns and locate faults, and finally a diagnostic report containing fault information is generated and output.

[0032] It can be understood that the fault positioning module uses this newly activated historical diagnosis model or newly generated new diagnosis model to perform anomaly detection on the real-time time series data stream; if an anomaly is detected, the abnormal features, such as a value of a certain sensor deviating from a predicted value by more than 20%, are input into the fault knowledge graph; the knowledge graph reasoning engine performs root cause tracing according to a preset correlation rule and outputs a number of reasons with the highest possibility; the system integrates all information to generate a diagnosis report containing a fault location, possible root causes, a severity level, and maintenance suggestions, and pushes the report to an operation and maintenance personnel; the process returns to step S01, and the system continues monitoring under the new diagnosis model, realizing self-adaptation to environmental changes.

[0033] It can be understood that data is the basis of the entire system, and high-quality and structured data input is a prerequisite for ensuring the accuracy of all subsequent analyses; the data collection module of the present application solves the problem of disorganized and difficult-to-use raw data through sensor network layout and building equipment layer models.

[0034] Specifically, in the embodiments of the present application, for example, in a large commercial complex, the deployment of the sensor network is hierarchical; for example, in a central refrigeration station, temperature, pressure, and liquid level sensors are installed on the evaporator, condenser, and compressor of the chiller unit; flow meters and pressure difference sensors are installed on the chilled water and cooling water circulation pipelines; current, voltage, and power factor sensors are installed in the power distribution cabinet of the unit; at the air conditioning box and fresh air unit, temperature and humidity sensors are installed at the inlet and outlet air inlets and outlets, and before and after the coil; wind speed or pressure difference sensors are installed in the air duct; state feedback and opening degree feedback sensors are installed on the fan and water valve actuators; at the end devices, such as the fan coil units distributed in each room, the supply and return air temperature, water valve opening, and fan gear are monitored; a large number of indoor temperature and humidity and carbon dioxide concentration sensors are deployed in the building space to reflect the actual environmental conditions.

[0035] These sensors come from different sources, such as different brands or different communication protocols, have different sampling frequencies, such as from seconds to minutes, and have non-uniform data formats. In order to manage these complex data, the present application introduces the concept of a building equipment layer model, which is not a physical entity but a logical data organization architecture that organizes all devices and their measurement points in the building in a tree or graph structure according to physical association and functional relationship. The building equipment layer model can be divided into a device layer, a subsystem layer, and a system layer.

[0036] Specifically, the device layer is the bottom layer of the building device layer model, representing a single independent physical device, such as a 1# chiller, an air handling unit located on the 3rd floor of the east zone, and a fan coil unit in the conference room. This layer is associated with the measurement point information of all sensors directly installed on the device, such as the evaporator outlet water temperature of the 1# chiller.

[0037] The subsystem layer is the middle layer of the building device layer model, representing a collection of multiple devices that work together to complete a specific function; for example, the refrigeration subsystem includes all chiller units, chilled water pumps, cooling towers, and related valves; the fresh air subsystem includes all fresh air handling units and their air duct networks; the data at this layer not only includes the data of all subordinate device layers, but also may include some macro indicators calculated, such as the total energy efficiency ratio of the refrigeration station.

[0038] The system layer is the highest layer of the building device layer model, representing the entire building's energy and environmental control system, integrating all subsystem data for global energy efficiency analysis and status monitoring; for example, the total power consumption and total water consumption of the entire building.

[0039] Through this building device layer model, the originally chaotic thousands of measurement points are given a clear contextual relationship, and when a fault occurs, the impact range can be quickly located along this hierarchical structure, whether it is from a single device to a subsystem or a problem within a subsystem.

[0040] It can be understood that on the basis of the building device layer model, the data integration and standardization processing process includes time alignment, missing value filling, and normalization / standardization.

[0041] Specifically, due to different sampling frequencies of different sensors, it is necessary to unify them to a common time reference, i.e. time alignment, for example, a time granularity of 5 minutes can be set, for data with a sampling frequency higher than 5 minutes, mean, maximum or minimum aggregation operations can be performed; for data with a sampling frequency lower than 5 minutes, forward or backward padding is performed. Due to network communication failure or sensor damage, missing values inevitably appear in the data stream, at this time, missing value filling is needed, according to the data characteristics, a variety of methods can be used for filling, for example, for temperature data with slow changes, linear interpolation can be used; for data with obvious periodicity, a filling method based on historical synchronous data can be used; or a machine learning model is used to predict missing values according to other related variables. The units and numerical ranges of different physical quantities are very different, for example, temperature values are between -10 and 40, while flow values can be between 0 and 1000 cubic meters / hour, in order to eliminate the influence of dimension and make different features have equal importance in model training, normalization (scaling data to the [0, 1] interval) or standardization (transforming data into a distribution with a mean of 0 and a standard deviation of 1) is needed.

[0042] After the above processing, the system finally obtains one or more structured time series data high-dimensional matrices, where each row represents a time step, and each column represents a feature corresponding to a building device layer model node.

[0043] It can be understood that the dual-time-scale concept drift detection module enables the system to both quickly respond to sudden events and remain sensitive to slow changes, thereby comprehensively understanding the dynamics of the data stream.

[0044] Specifically, the detection mechanism of a single time scale has inherent contradictions. If the window is short, it will be very sensitive to sudden failures, such as the sudden stop of a water pump, but it will be easily disturbed by short-term noise and will not be able to detect slow drifts that need to be accumulated for a long time, such as device aging; if the window is long, it will have good detection effect and strong stability for slow drifts, but it will be very slow in response to sudden failures, losing the meaning of early warning.

[0045] The dual-time-scale design is to solve this contradiction. The short-term sliding window is always concerned about the instantaneous data fluctuation, while the long-term baseline window examines whether the current trend is reasonable by comparing it with the distant past. The two work in parallel and complement each other.

[0046] Short-term mutation drift detection is to quickly detect rapid changes in data distribution caused by sudden equipment failure, sudden change in external environment or human error.

[0047] For example, cumulative sum control charts in statistical process control are very sensitive to small, sustained shifts in process mean.

[0048] Take the running current of a fan with constant load as an example. Normally, the current value should fluctuate within a small range. Suppose the normal running current of the fan is 50 amperes, and the standard deviation is 1 amperes.

[0049] First, set the target value (T), for example, 50 amperes, and the allowed offset (K), for example, 0.5 times the standard deviation, i.e. 0.5 amperes.

[0050] Then, define two cumulative sums: one for detecting upward shifts in mean (SH) and one for detecting downward shifts in mean (SL).

[0051] For each newly collected current value x(t), update the cumulative sums as follows: SH(t) = max(0, SH(t-1) + (x(t) - (T + K))); SL(t) = min(0, SL(t-1) + (x(t) - (T - K))); Set the control limit (H), for example, 5 times the standard deviation, i.e. 5 amperes.

[0052] Suppose the fan bearing starts to jam, causing the motor load to increase slightly but continuously, and the current value becomes 51 amperes, 51.2 amperes, 51.1 amperes, etc. Although the single measurement value is still within the normal fluctuation range, the cumulative sum control chart algorithm will accumulate these small positive deviations.

[0053] The value of SH will continue to grow, and when SH(t) exceeds the control limit H (5 amperes), the system will immediately issue a sudden shift alert.

[0054] The length of the short-term sliding window is crucial. For building systems, it is usually set to the past 1 to 6 hours of data, which is sufficient to cover most operating shifts and immediate response periods. The settings of control limits K and H need to be adjusted empirically or statistically calibrated based on historical data according to the stability of the specific monitored variable and the required detection sensitivity.

[0055] Long-term gradual and recurring shift detection is to identify patterns that are not obvious in a short period of time, but accumulate or periodically occur over weeks, months or even a year.

[0056] Specifically, for long-term shifts, the test method based on data distribution difference is more effective, among which the K-S test (Kolmogorov-Smirnov Test) is a commonly used non-parametric test method for comparing the difference between the empirical cumulative distribution functions of two samples.

[0057] Take the example of detecting the gradual concept shift of a central air conditioning system from spring to summer; define a current window, e.g. containing all data from the past week; at the same time, define a baseline window, whose data comes from an earlier period that is confirmed to be stable, e.g. a week one month ago, or a week from the same period last year.

[0058] Select several key variables that can reflect the core state of the system, e.g. chilled water supply and return temperature difference, outdoor temperature, total building cooling load; for each key variable, calculate the empirical cumulative distribution function of the data in the current window and the baseline window respectively; the core of K-S test is to find the maximum absolute difference between the two cumulative distribution functions at all points, i.e. the D statistic; compare the calculated D statistic with the critical value, which depends on the sample size and the set significance level (e.g. α = 0.05); if the D statistic is greater than the critical value, the null hypothesis that the two samples come from the same distribution can be rejected, thus determining that a significant gradual concept shift has occurred.

[0059] For example, in spring, the chilled water temperature difference may be close to 0 for a long time, and the total building cooling load is also low; when entering early summer, as the outdoor temperature rises, the chiller starts to start, the value distribution of chilled water temperature difference and total cooling load will shift to the right as a whole, and the shape of its cumulative distribution function curve will change significantly; at this time, the D statistic calculated by K-S test will increase significantly and exceed the threshold, and the system determines that a gradual drift has occurred.

[0060] When a long-term drift is detected, the system needs to determine whether it is a completely new concept or a recurrence of a historical concept.

[0061] First, extract the macro feature vector of the current window data, which can include the mean, variance, quantile of each key variable, and the correlation coefficient matrix between variables, etc.

[0062] Then, the system can traverse the metadata tags of all diagnostic models in the dynamic model pool, which may also contain a similar historical feature vector.

[0063] By calculating the similarity between the current feature vector and each historical feature vector, e.g. using cosine similarity or Euclidean distance, the most matching historical pattern can be found.

[0064] If the highest similarity score exceeds a pre-set threshold, e.g. 0.9, the system determines that this drift is a historical concept recurrence, and determines which specific historical concept it is; otherwise, it is determined to be a new concept.

[0065] It can be understood that the model management module can intelligently schedule and maintain the diagnostic model according to the result of drift detection. The dynamic model pool is a living knowledge base that stores and manages all valuable experiences accumulated by the system during long-term operation. Each diagnostic model in the dynamic model pool can be constructed by selecting a suitable algorithm according to the different characteristics of the diagnostic object.

[0066] Specifically, the two types of recurrent neural networks, LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit), are particularly good at processing time series data. For diagnostic tasks with obvious seasonality, periodicity or time dependence, such as predicting future energy consumption for one hour and comparing it with the actual value to find abnormalities, LSTM and GRU are ideal choices. For example, an LSTM-based model can learn the completely different energy consumption patterns of weekdays and weekends, and day and night.

[0067] The autoencoder is an unsupervised deep learning model that is very suitable for anomaly detection. Its working principle is to compress the input data into a low-dimensional latent representation through a neural network, and then try to reconstruct the original input from this latent representation. After training only with normal operation data, this deep learning model can reconstruct normal data well, and the reconstruction error is small. When inputting abnormal data, for example, a sensor reading deviates from the normal mode due to a fault, the deep learning model will not be able to reconstruct it well, resulting in a significant increase in reconstruction error, and thus being identified as an anomaly.

[0068] It can be understood that when the incremental learning unit successfully generates a new diagnostic model for a new concept with stable performance, the new diagnostic model will be encapsulated and stored in the dynamic model pool; at the same time, the system will automatically or semi-automatically generate metadata tags for it; for example, the metadata tags can include "generation time: August 2025", "triggering event: first encounter with continuous 40-degree extreme high temperature weather", and "applicable data features: outdoor average temperature > 38°C, full load operation". In order to prevent the model pool from expanding indefinitely, an elimination mechanism is needed; for example, some very old and long-inactivated models, or models whose applicable range is completely covered by updated or superior diagnostic models, can be archived or deleted. In addition, the diagnostic model itself is static, but its metadata can be updated; for example, as the understanding of a certain operating mode deepens, the metadata tags can be manually optimized to be more accurate, so that they can be more accurately matched in the future.

[0069] It can be understood that when the system encounters a new concept that has never been encountered before, the incremental learning unit is activated, which can quickly learn new knowledge without completely discarding existing knowledge.

[0070] Specifically, assume the baseline model the system is running is Model A, which is good at handling data of "spring ventilation mode". Now, the system detects a brand new "ultra-long plum rain season mode" (i.e. a new concept), whose temperature-humidity coupling relationship is never seen by Model A.

[0071] The concept drift detection module sends a new concept signal, and the model management module decides to start incremental learning; the system will not directly train on the new data on Model A, because it is likely to destroy the weight parameters Model A has learned about the spring mode, i.e. catastrophic forgetting; the system will create a copy of Model A, or a new model with the same structure but randomly initialized parameters, called Model B.

[0072] Next, let Model B use the newly collected "plum rain season" data for training. For example, the prediction loss is the standard training loss, which requires the prediction results of Model B to be as close as possible to the true labels of the "plum rain season" data; at the same time, the system will use a part of the old "spring" data (or generated data that can represent the old mode) to test Model B, and require the output of Model B (not only the final prediction result, but also the intermediate layer activation value or output probability distribution) to be as consistent as possible with the output of Model A on the same data, i.e. the knowledge preservation loss.

[0073] By optimizing this compound loss function, Model B will eventually become a new model that can handle the "plum rain season" new mode well, and to a large extent, retain the ability to handle the "spring" old mode; after training is complete, this more comprehensive performance Model B will replace Model A and become the system's current latest diagnostic model. In this way, the system achieves efficient iteration of knowledge and avoids the huge overhead of starting from scratch every time a new situation is encountered.

[0074] It can be understood that detecting an anomaly is only the first step, and more importantly, it tells the operation and maintenance personnel where the problem is and what the possible cause is. The fault locating module of the present application introduces a fault knowledge graph.

[0075] The fault knowledge graph is a structured semantic network used to describe the knowledge in the field of fault diagnosis, composed of nodes and edges. Nodes can represent various entities, such as device components, sensors / measuring points, fault phenomena / abnormal patterns, fault causes, or maintenance operations. Device components can be compressors, condenser fans, expansion valves; sensors / measuring points can be compressor discharge pressure, evaporator outlet water temperature; fault phenomena / abnormal patterns can be high discharge pressure, insufficient refrigeration capacity, abnormal increase in energy consumption; fault causes can be excessive refrigerant, poor condenser heat dissipation, filter blockage; maintenance operations can be checking condenser coils, supplementing refrigerant. Edges represent the relationship between nodes, such as: a water chiller includes a compressor, a discharge pressure sensor is used to measure the discharge pressure of the compressor, excessive refrigerant is manifested as high discharge pressure, high discharge pressure is caused by excessive refrigerant, and excessive refrigerant recommends recovering part of the refrigerant.

[0076] This fault knowledge graph can be constructed and continuously expanded through various channels such as the experience of field experts, device maintenance manuals, and historical maintenance records.

[0077] It can be understood that when the diagnostic model detects one or more abnormal measuring points, i.e., the reconstruction error exceeds the threshold, the positioning process is started: the system converts the abnormal measuring points and their deviation from the normal pattern into fault phenomenon nodes in the graph.

[0078] For example, the evaporator outlet water temperature is consistently 5°C higher than the predicted value, and the compressor current is 20% higher than the normal pattern. The system starts with these fault phenomenon nodes and performs a reverse query in the knowledge graph, tracing along the "caused by" relationship edges. If there is only one abnormal phenomenon, such as high discharge pressure, the fault knowledge graph may return multiple possible root causes, such as excessive refrigerant, poor condenser heat dissipation, and non-condensable gas in the system. If multiple abnormal phenomena are detected simultaneously, such as high discharge pressure and excessive supercooling, the possible causes of the two phenomena are intersected or weighted scored; because poor condenser heat dissipation usually leads to a decrease in supercooling, while excessive refrigerant only causes both high discharge pressure and excessive supercooling; therefore, by fusing multiple evidences, the confidence of excessive refrigerant will be significantly improved. Finally, output a list of root causes sorted by possibility from high to low; for example: [(excessive refrigerant, 95% confidence), (poor condenser heat dissipation, 10% confidence)].

[0079] Finally, the system generates a diagnostic report that clearly presents: the No. 2 water chiller has failed, with the abnormal performance of high discharge pressure; the most likely root cause is excessive refrigerant; the recommended maintenance operation is to check the refrigerant charge and perform appropriate recovery; the fault severity level is moderate.

[0080] In other embodiments of the application, for example, a large office building, the time is the end of May, the temperature gradually rises, the central air conditioning system is in the transition from the spring pure ventilation mode to the summer cooling mode.

[0081] The system is currently running a diagnostic model "Spring Ventilation Model", which is trained on the data in March, and its knowledge base mainly contains the relationship between fan operation and indoor and outdoor temperature and humidity changes, and almost no knowledge about the operation of the chiller unit.

[0082] As the outdoor temperature exceeds 26°C for several consecutive days, multiple areas in the building begin to request cooling through the building automation system, and the property management personnel start the 1st and 2nd chiller units of the central cooling station. The total energy consumption curve of the system, the chilled water supply and return water temperature and other key indicators begin to show a completely different form from the pure ventilation mode.

[0083] The long-term baseline window of the concept drift detection module starts to work, for example, the data of the past week is the current window, and the data of a week ago is the baseline window. Through K-S test on the feature "building total power consumption", it is found that the data distribution of the current window (showing a high peak in the daytime due to cooling) and the data distribution of the baseline window (relatively flat) are significantly different, and the D statistic far exceeds the critical value.

[0084] The system determines that a slow concept drift has occurred, and extracts the macro feature vector of the current week's data, such as daily average power consumption, power consumption peak-valley difference, and chilled water average temperature difference. The feature vector is matched with the metadata of all historical models stored in the dynamic model pool.

[0085] The matching result shows that the current feature vector has a similarity of 0.92 with a historical model named "Early Summer Transition Season Model". The metadata of this model indicates that it was generated at the beginning of last June.

[0086] Therefore, it loads the "Early Summer Transition Season Model" into the memory from the model pool and activates it as the current dominant fault diagnosis model. This new model is an autoencoder-based model that understands what the parameters of the chiller unit should be under the current outdoor temperature and indoor load.

[0087] After the system has been smoothly switched to the "early summer transition season model" and runs for several days, a very slight refrigerant leakage occurs at a pipe joint of the No. 2 chiller unit, with a very slow leakage rate that does not cause an obvious pressure drop or other alarms in a short time. Due to the lack of refrigerant, the refrigeration efficiency of the No. 2 unit begins to decline, and in order to achieve the set chilled water outlet temperature, for example, 7°C, the control system of the unit has to make the compressor run at a higher load rate or start and stop more frequently. This is reflected in the data that the "percentage of compressor running time of the No. 2 unit" feature value begins to be consistently and slightly higher than the normal value predicted by the model under the current operating conditions; at the same time, the "chilled water supply and return water temperature difference" does not increase accordingly, but decreases slightly.

[0088] The real-time sensor data stream is sent to the activated "early summer transition season model", which reconstructs the "normal" data according to the input of dozens of related variables such as outdoor temperature, cooling water temperature, load request, etc. When calculating the reconstruction error of the two dimensions of "percentage of compressor running time of the No. 2 unit" and "chilled water supply and return water temperature difference", the system finds that their error values begin to rise consistently and slowly; although the single error value may not be large, the anomaly detection algorithm in the short-term sliding window finds that, for example, the reconstruction error at consecutive time points is smoothed and averaged, and the average error has stably exceeded the preset anomaly threshold; the system determines that the No. 2 chiller unit is abnormal.

[0089] The fault location module receives the anomaly signal from the diagnostic model and the specific abnormal features: "abnormally high compressor load rate" and "decline in refrigeration efficiency"; taking these two "fault phenomenon" nodes as input, it queries the internal fault knowledge graph.

[0090] The graph reasoning engine begins to trace back: "decline in refrigeration efficiency" can be caused by "lack of refrigerant", "poor heat dissipation of condenser", "low evaporator heat exchange efficiency", etc.; "abnormally high compressor load rate" also points to these causes.

[0091] In order to further distinguish, the reasoning engine queries the status of other sensors related to these possible causes, checks the "cooling tower fan current" and "cooling water outlet temperature" related to "poor heat dissipation of condenser", and finds that they are both within the normal range predicted by the model, and also checks the "chilled water flow" related to "low evaporator heat exchange efficiency", and finds that it is also normal.

[0092] Through elimination, the graph reasoning engine greatly improves the confidence of "lack of refrigerant" as the root cause, and the system finally generates a high-priority alarm and pushes it to the mobile APP of the operation supervisor.

[0093] The report content is: "Alert: No. 2 chiller suspected failure; failure phenomenon: under the current load, the compressor running time exceeds the normal range by 15%, but the refrigerating capacity is not up to standard; root cause inference: refrigerant leakage, confidence 90%; recommended operation: immediately send technicians to carry pressure gauges and leak detectors to check the refrigerant system of No. 2 unit; evaluate the severity: moderate, if not handled, it may cause compressor damage and sharp rise in energy consumption."

[0094] Through this complete example, it is shown how to seamlessly integrate various modules, from perceiving macro changes in the environment, to intelligently scheduling historical knowledge, to accurately capturing micro failure signals, and finally providing decision support with deep insights, thus forming a truly self-adapting and continuously learning intelligent fault diagnosis system.

[0095] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional modules is exemplified, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0096] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the modules or units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0097] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0098] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The above integrated unit can be realized in the form of hardware or in the form of software functional unit.

[0099] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory, random access memory, magnetic disk or optical disk, and various media that can store program codes.

[0100] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. Any feature disclosed in the specification (including the abstract and drawings) can be replaced by other equivalent or similar features, unless specifically stated. That is, each feature is only an example of a series of equivalent or similar features, unless specifically stated.

Claims

1. A method for locating a fault of a building equipment based on a sensor network and a layer model, characterized by, The method comprises the following steps: Data acquisition and preprocessing: continuously collecting multi-source and heterogeneous time series data from a sensor network covering building equipment, and integrating and standardizing the time series data based on a pre-set building equipment layer model; Dual-time scale concept drift detection: a short-term sliding window for detecting sudden drift of data flow and a long-term baseline window for detecting slow or recurring drift of data flow are set in parallel, and by continuously monitoring the data feature distribution in the short-term sliding window and the long-term baseline window, the concept drift events occurring in the building equipment operation state are identified and their types are determined; Dynamic model management and adaptive correction: a dynamic model pool containing multiple historical diagnostic models corresponding to historical concepts is constructed and maintained; when the concept drift event is detected, the following adaptive correction logic is executed: if the concept drift event is identified as a recurring historical concept, the corresponding historical diagnostic model is retrieved from the dynamic model pool and activated as the current diagnostic model; if the concept drift event is identified as a new concept, an incremental learning mechanism is triggered to generate a new diagnostic model online based on the existing model knowledge; Fault location and output: the historical diagnostic model or the new diagnostic model is used to analyze the real-time collected time series data to identify abnormal patterns and locate faults, and finally a diagnostic report containing fault information is generated and output.

2. The method of claim 1, wherein, The short-term sliding window is used to capture rapid data distribution changes caused by sudden equipment failure or sharp changes in operation mode, and the long-term baseline window is used to capture slow data distribution changes caused by seasonal changes, equipment aging or gradual changes in operation load.

3. The method of claim 2, wherein, In the short-term sliding window, statistical process control methods are used to monitor the instantaneous changes of key performance indicators to detect sudden drift; in the long-term baseline window, data distribution difference test methods are used to detect slow drift by comparing the overall data distribution difference between the current window and the historical window.

4. The method of claim 1, wherein, Each historical diagnostic model in the dynamic model pool is associated with a metadata tag describing the historical concept it corresponds to, which includes but is not limited to applicable season, working condition mode, equipment configuration or timestamp information.

5. The method of claim 4, wherein, The concept drift event is identified as a recurring historical concept, which includes: extracting the feature vector of the current time series data, and calculating the similarity between the feature vector and the historical concept features described by the metadata tags of the historical diagnostic models in the dynamic model pool; when the similarity exceeds a pre-set threshold, it is determined as a recurring historical concept.

6. The method of claim 1, wherein, The incremental learning mechanism uses knowledge distillation or parameter fine-tuning techniques to train the diagnostic model using the time series data corresponding to the new concept by constructing a composite loss function containing prediction loss and knowledge preservation loss, thereby realizing online updating without complete retraining to avoid catastrophic forgetting of historical knowledge.

7. The method of claim 1, wherein, The preset building equipment layer model is used to integrate and standardize the time series data, including mapping the collected sensor data to corresponding nodes in the equipment layer, subsystem layer and system layer, and performing time alignment, missing value filling and normalization operation on the time series data from different sources to form a unified format feature input.

8. The method of claim 1, wherein, The fault locating and outputting step further includes: matching and reasoning the abnormal pattern or abnormal score output by the historical diagnosis model or the new diagnosis model with the pre-constructed fault knowledge graph to determine the root cause of the fault.

9. The method of claim 8, wherein, The diagnosis report includes the equipment location where the fault occurs, the most likely root cause, the evaluated fault severity level and the recommended maintenance operation suggestion.

10. A sensor network and layer model based building equipment fault location system, characterized in that, It comprises: a data acquisition module configured to continuously acquire multi-source heterogeneous time series data from a sensor network covering building equipment, and integrate and standardize the time series data based on a preset building equipment layer model; a concept drift detection module configured to set a short-term sliding window for detecting sudden drift of data flow and a long-term baseline window for detecting slow drift or recurring drift of data flow in parallel, and identify concept drift events occurring in the building equipment operation state and determine their types by continuously monitoring the data feature distribution in the short-term sliding window and long-term baseline window; a model management module configured to build and maintain a dynamic model pool containing a plurality of historical diagnosis models corresponding to historical concepts respectively; when the concept drift event is detected, the following adaptive correction logic is executed: if the concept drift event is identified as a recurring historical concept, the corresponding historical diagnosis model is retrieved from the dynamic model pool and activated as the current diagnosis model; if the concept drift event is identified as a new concept, an incremental learning mechanism is triggered to generate a new diagnosis model online based on the existing model knowledge; a fault locating module configured to analyze the real-time collected time series data using the historical diagnosis model or the new diagnosis model to identify abnormal patterns and locate faults, and finally generate and output a diagnosis report containing fault information.

Images