WO2022165792A1

WO2022165792A1 - Method and system of sensor fault management

Info

Publication number: WO2022165792A1
Application number: PCT/CN2021/075763
Authority: WO
Inventors: Ka Kui LO; Sin Ying LEUNG; Guangya Zhu; Xiaojun Luo; Yongjun SUN; Yau Choi NGAI; Kwok Hi LEUNG; Seung Hyo BAEK; Kwong Fai Fong; Chun Kwong Lee
Original assignee: Siemens Limited, Hong Kong
Priority date: 2021-02-07
Filing date: 2021-02-07
Publication date: 2022-08-11

Abstract

A method and a system of sensor fault management for a chiller plant is disclosed. The method comprises the steps of: obtaining (S310) a fault-free dataset of sensors for the chiller plant; generating (S320) a faulty dataset by introducing instantaneous sensor signal errors into the fault-free dataset; running (S330) a system model with the fault-free dataset and the faulty dataset as inputs to obtain a modeled dataset, wherein the system model is set up by deep learning with operating parameters and the fault-free dataset; building up (S340) a fault pattern database including fault-free profiles and faulty profiles based on the modeled dataset; evaluating and reconstructing (S350) sensor data for evaluation with reference to the fault pattern database. This is a smart sensor fault detection and diagnosis solution through big data analytics and it is applicable for multiple sensors for various chiller plants.

Description

A Method and a System of Sensor Fault Management

Technical Field

The present disclosure relates to a method of Sensor Fault Management (SFM) for Heating and/or Ventilation and/or Air-Conditioning (HVAC) . More specifically, the present disclosure substantially limits to the SFM application in a chiller plant that is installed for a residential and/or commercial and/or industrial site.

Background

HVAC systems regulate the comfort levels of many indoor environments. A chiller plant is a centralized system that cools the air for a building or for a collection of buildings and provides the air-conditioning portion of HVAC systems. A chiller plant is an energy-consuming system, as around 40 percent of the energy consumed by buildings is used for the chiller plant. One way to save energy in chiller plants is through operations and maintenance. Today, many systems connect to centralized controls, e.g. Building Management System (BMS) , that can automatically adjust output to maintain the most efficient operations of a chiller plant. Keeping mechanical units functioning at full capacity requires periodic maintenance, such as changing/fixing those fault devices in time upon fault diagnosis.

The operation of a chiller plant is affected by the changing ambient air and loading conditions, as well as the reliability of the equipment members and control components. In the chiller plant, therefore, reliable and accurate sensor measurements have become essential for monitoring the whole system performance, implementing control strategies and diagnosing equipment and system performance. Since different kinds of sensor faults may occur in a chiller plant, SFM is useful and somehow critical to ensure the proper system operation.

Summary

The present disclosure teaches a method and a system of sensor fault management for a chiller plant. The method comprises the steps of obtaining a fault-free dataset of sensors for the chiller plant; generating a faulty dataset by introducing instantaneous sensor signal errors into the fault-free dataset; running a system model with the fault-free dataset and the faulty dataset as inputs to obtain a modeled dataset, wherein the system model is set up by deep learning with operating parameters and the fault-free dataset to characterize performance of the chiller plant; building up a fault pattern database including fault-free profiles and faulty profiles based on the modeled dataset, where a profile of the fault-free profiles or the faulty profiles refers to one or more feature measurements characterizing fault-free cases or faulty cases; evaluating and reconstructing sensor data for evaluation, wherein the sensor data for evaluation is identified as fault and reconstructed when a faulty profile fits with the sensor data.

The method of sensor fault management employs deep learning algorithm to model the chiller plant performance, builds up a fault pattern database including both fault-free profiles and faulty profiles based on the modeled dataset, and identify fault (s) when the sensor data for evaluation fits well with a faulty profile. With this method, the fault sensor data may be identified automatically without a manual check, and even corrected/reconstructed in time by means of the fault pattern database, without replacing the fault sensors immediately. This provides BMS with more reliable inputs and thus ensuring a more robust chiller plant from control and management aspect.

In one example, the fault-free dataset preferably includes the data collected only from those critical sensors, to which the performance of the chiller plant is more sensitive than other sensors under fault situation. For example, a sensor may be identified as the critical sensor if a change of the performance of the chiller plant over an entire data range of the sensor is greater than a threshold, while other sensors are being kept at their median value. When only the critical sensors are considered, the deep learning algorithm with fault-free dataset may have lower computational complexity and higher efficiency.

In another example, the operating parameters and fault-free data for deep learning comprise: the fault-free data of sensors in current time step and those fault-free data of sensors in previous time step; preferably, the operating parameters and fault-free data at input layer of the deep learning are any of or any combination of followings: outdoor temperature, temperature of chilled water at return header, temperature of chilled water at supply header, temperature of cooling water at return header, temperature of cooling water at supply header, mass flow rate of chilled water, mass flow rate of cooling water, or mass flow rate of cooling tower air; preferably, the operating parameters and fault-free data at output layer of the system model are any of or any combination of followings: mass flow rate of chilled water, mass flow rate of cooling water, mass flow rate of cooling tower air, power consumption of chilled water pump, power consumption of cooling tower, power consumption of cooling power pump, number of chillers, number of chilled water pumps, numbers of cooling towers, number of cooling power pumps, temperature of chilled water at return header, or temperature of chilled water at supply header.

In another example, data collected from sensors are refined to form the fault-free dataset by implementing any of the steps of: converting non-numeric data to numeric data; checking for missing data, checking for data out-of-range, or checking for data conflict with equipment on/off status.

During the data refinement procedure, the quality of the logged data from BMS systems may be checked, which is useful as some information may be collected regarding the health of the sensors through this process. With the data refinement, the incomplete data, the conflict data or the data out of a reasonable range may be identified at an early stage, and thus the refined plant data are more reliable. This would be beneficial to forming an accurate system model from the starting point.

In a further example, to build the faulty database, different fault cases are considered. That is, the instantaneous signal errors may comprise any of followings:

- instantaneous signal error for bias, which is a positive or negative constant throughout a whole evaluation period;

- instantaneous signal error for drift, the magnitude of which increases or decreases with time;

- instantaneous signal error for precision degradation, which is determined from a Gaussian distribution function with mean value zero and standard deviation.

In a preferable example, the step of evaluating and reconstructing sensor data for evaluation comprises: obtaining a plurality of evaluation profiles over an evaluation period, each evaluation profile is based on signal readings of one or more sensors selected for evaluation at a predetermined evaluation time interval, the predetermined evaluation time interval is same as the time interval for fault pattern database; comparing the evaluation profile for each evaluation time interval with all of the faulty-free profiles and faulty profiles of the fault pattern database one by one; for each evaluation interval, identifying a fault profile that fits well with the evaluation profile, taking a fault case with highest frequency of occurrence throughout the whole evaluation period as a fault diagnosis result for the selected sensors.

In an example, the data-mining method is a clustering algorithm to group the modeled data into clusters and the profile comprises a feature measurement describing characteristic of each cluster. Preferably, the clustering algorithm is k-means clustering algorithm and the profile comprises a total centroid score at any time step i , CS _t, i, given by

where

X _i (i = 1, 2, …t _max) is a normalized sensor signal, t _max is the maximum time step in a predetermined period for the data-mining,

c _k (k = 1, 2, …N) is a centroid of the kth cluster for the fault-free cases, N is the number of clusters and given by a rule-of-thumb;

G _k, ff (k = 1, 2, …N) is the kth clustered data subsets for the fault-free cases;

n _k is the number of time steps in the kth clustered subset G _k, ff containing that at time step i.

In an example, a fault profile fits well with the sensor data when having a smallest Euclidean distance between the faulty profile CS _t, i, ft, and the evaluation profile of the sensor data CS _t, i, e. With the centroid score CS _t, i, a normalized parameter, the system modeling and evaluation are more reliable.

In one example, the method further comprises: with the reconstructed sensor data, calculating bias for non-critical sensor at different locations of the chiller plant; generating trend data of the bias for the non-critical sensors; determining a representative bias throughout a whole record period for the non-critical sensors by flattening the trend data. In another example, the step of flattening the trend data comprises: segmenting the trend data into various groups by using a moving average approach; leveling the trend data of each group based on a group average value; merging of adjacent groups with group average values differed by less than a predetermined threshold; recalculating a new average value for each merged group and repeating the merging until no more group merging.

According to one aspect of the present disclosure, it is disclosed a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of any of the methods as mentioned above. It is also disclosed a computer-readable medium having stored thereon the computer program as mentioned above.

According to another aspect of the present disclosure, it is disclosed a system for sensor fault management for a chiller plant, comprising a computing device configured to communicate with a Building Management System that controls a chiller plant installed with a plurality of sensors. The computing device is configured to implement the above methods. For example, the computing device may be configured to:

receive a fault-free dataset of the sensors;

generate a faulty dataset by introducing instantaneous sensor signal errors into the fault-free dataset;

run a system model with the fault-free dataset and the faulty dataset as inputs respectively to obtain a modeled dataset, wherein the system model is set up by deep learning with operating parameters and the fault-free data to characterize the performance of the chiller plant;

build up a fault pattern database including fault-free profiles and faulty profiles by applying a data-mining method to the modeled dataset, where the profile refers to one or more feature measurements characterizing fault-free cases or faulty cases;

evaluate and reconstruct sensor data for evaluation, wherein the sensor data for evaluation is identified as fault and reconstructed when a faulty profile fits with the sensor data.

A novel SFM and sensor data reconstruction solution for chiller plants through big data analytics is provided in present disclosure. The most preferable solution is devised which employ deep learning algorithm to model the chiller plant performance and clustering method to predict various types of sensor faults for critical sensors; while thermodynamics approach to determine the bias for non-critical sensors. A full summary report for fault reconstruction of both critical and non-critical sensors, as well as trend plots for non-critical sensors, could be generated for the maintenance staff to take any remedial actions. Various chiller plant logged data as well as site checking reports are employed to validate the solution. The proposed SFM solution may be considered as not only a robust choice, but also a foundation for future because of its higher accuracy and reliability and less manpower.

Brief description of the drawings

Various features will become apparent to those skilled in the art from the following detailed description of the disclosed non-limiting examples. The drawings that accompany the detailed description can be briefly described as follows:

FIG 1 is a schematic drawing of a typical chiller plant.

FIG 2 is a schematic drawing of a SFM system according to an example of the present disclosure.

FIG 3 shows an overview flow chart of SFM method according to an example of the present disclosure.

FIG 4 shows a flow chart of a pre-handling method for the SFM method of FIG 3 according to an example of the present disclosure.

FIG 5 shows a flow chart of an example of SFM method, in which an evaluation and reconstruction solution is illustrated according to an example of the present disclosure.

FIG 6 shows a flow chart of an example of SFM method, in which sensor signal bias is identified and reconstructed for those non-critical sensors according to an example of the present disclosure.

FIG 7 shows an example of temperature bias trend plots for non-critical sensors.

Detailed description

References are made to the accompanying drawings that form a part hereof, and in which is shown by way of illustration various examples. It is to be understood that the terms used herein are for the purposes of describing the figures and examples and should not be regarded as limited in scope.

FIG 1 shows a typical water-cooled chiller plant 100, which comprises chillers 112, chilled water pumps 116, cooling towers 124 and cooling water pumps 126.

FIG 1 shows a chiller plant 100 including at least three chillers 112. It will be obvious that the concepts described herein may be implemented in any chiller plants with one or more chillers with the appropriate fluid line connections. If there are at least two chillers, each chiller may run independently at a different temperature. Chilled water circuits of the chillers may form fluid communication when one of the chillers may fail or be out of service, which allows the remaining operable chiller (s) to provide backup operation to the failed chiller (s) , as shown in FIG 1. The total number of chillers in a chiller plant depends on the specific requirements and/or restriction.

The chilled water pumps 116 pump chilled water out to the building 130 through a chilled water circuit 110 to e.g. air handling units at the building side, where the air handling units use valves to vary the amount of chilled water to that unit's water coil to control the temperature of the air coming out of the unit. Thus, the added heat comes back into the water that goes along the chilled water circuit 110 back to the chiller plant. The term “chilled water circuit” generally refers to a chilled water circuit and the evaporator of the connected chiller.

The return water from the building 130 goes to the evaporator of the chillers and transfers its heat to the condenser of the chillers, which is in a separate cooling water circuit 120.

The cooling water pumps 126 pump the cooling water out to the cooling towers 124, which are sometimes outside. The water is pumped to the top of the cooling tower 124 and then rains down inside it. There are big fans on the top of the cooling tower 124 that pulls air through the tower, thus across the raining water and then blows it outside. This water pools up at the bottom of the cooling tower 124 and is sucked back and run through the condenser side of the chillers again along the cooling water circuit 120, thus removing the heat from the building 130. The term “cooling water circuit” generally refers to the cooling water circuit and the condenser side of the connected chiller.

There may be multiple sensors 140 allocated in the chiller plant 100, as illustrated in FIG 1. These sensors 140 may be an independent sensor, e.g. a water flow sensor or a temperature sensor, or sensors integrated with equipment, e.g. power consumption reading. Generally, in the following description, “sensor” refers to not only the physical sensors installed in the chiller plant, but also the field components or equipment installed where various measurements may be read out.

FIG 1 only shows an example of a water-cooled chiller plant. In centralized building air-conditioning systems, various chiller plants are designed, the most common types of which are:

· Multiple air-cooled chillers using differential pressure bypass (DPB) or de-coupler or variable primary flow for chilled water supply; and

· Multiple water-cooled chillers using cooling towers for heat rejection and DPB/de-coupler or variable primary flow for chilled water supply.

Studies on SFM of air-conditioning systems have been made in recent decades. They are focused mainly on individual piece of equipment only or based on the specific design of a chiller plant. That is, the existing solutions require full understanding of the design and control strategy applied in a specific chiller plant, which may differ significantly from chiller plants to chiller plants in actual practice. Therefore, the inventors of the present disclosure propose a new SFM solution, which is easily adaptive to different kinds of chiller plants.

FIG 2 shows a control system 20 for managing a chiller plant according to an example of the present disclosure. In the example, the system 20 includes a computing device 200, which is configured to communicate with a BMS 280. BMS is typically used to control and monitor the installed equipment including mechanical and electrical equipment and systems like HVAC system. The BMS 280 may be implemented on a stand-alone computer, a local computing server or a cloud-based server. The BMS may communicate with sensors 140 or other equipment, like chillers 112 or

pumps

116 and 126, arranged in the chiller plant 100, so as to exchange data (all parameters or sensor signal readings) or commands with them. The computing device 200 may exchange data or commands with BMS 280 via wired or wireless communication. Alternatively, the computing device 200 may be implemented on one device of the BMS. That is, the control system 20 itself is a subsystem of the BMS.

As shown in FIG 2, the computing device 200 includes a processor 210, a memory 220 and a communication module 230 for data exchange with or data collection from BMS 280. The computing device 200 may be also a computing device arranged in the same building as the chiller plant, or a remote server allocated in another building, or even a cloud-based server. In the example where the computing device 200 is implemented on a cloud-based server of BMS, sensor signal readings, which are collected and stored as logged data in BMS, are transferred from BMS 280 to the computing device 200 via a wire or wireless communication channel by means of the communication module 230. Computer programs that the processor 210 executes to implement the SFM may be stored in the memory 220 together with various database/datasets needed. The memory 220 may be divided into several sub-blocks 221～225 for different datasets. The details of memory 220 will be described later with reference to FIGs 3-6.

As shown in FIGs 1 and 2, the communication module 230 is communicative with, and able to receive logged data from BMS 280. BMS 280 is able to get sensor signal readings from the sensors 140 or other installed equipment, like chillers 112, cooling towers 122, pumps 114 and 126. For example, in the example of FIG 1, the sensor 140 may include various sensors of the chiller plant 100. Some of the sensors 140 may be for example temperature sensors that record a signal indicative of a temperature of chilled water at chilled water supply header; a signal indicative of a temperature of chilled water at chilled water return header; a signal indicative of a temperature of cooling water at cooling water return header, or a signal indicative of a temperature of cooling water at cooling water supply header. Some of the sensors 140 may also be mass flow rate sensors (e.g. flow rate meter) that record a signal indicative of the mass flow rate of the chilled water circuit 110, or a signal indicative of the mass flow rate of the cooling water circuit 120. In an example, the flow rate meter is or comprises an ultrasonic flow meter. In another example, the flow rate meter employs a physical/mathematic algorithm to calculate out a mass flow rate of a circuit.

In FIG 2, the computing device 200 may receive sensor signal readings from BMS 280, and then implement the SFM method provided here according to an example of the present disclosure to identify the faulty sensor signals, preferably with deviation values, or even more preferably, further reconstruct the sensor signal reading which is identified as fault. The identified faulty sensor signals, deviation values, or even the reconstructed sensor signals may be used to optimize the control strategy or maintenance strategy of the chiller plant at the BMS side.

FIG 3 shows an overview flow chart of the SFM method for a Chiller Plant according to an example of the present disclosure. In the example of FIG 3, the SFM method is implemented by the computing device 200. In other words, the SFM method may be implemented as a software platform allocated on the computing device 200.

As shown in FIG 3, in Step S310, the computing device 200 obtain a fault-free dataset of sensors. In one example, the fault-free dataset may be set up by collecting data/signal readings from sensors immediately after sensors’ calibration, and the fault-free dataset may be stored in the memory block 221a. In another example, the sensor signals being read within the first several months (e.g. 6 months) after installation are taken as the fault-free data and used to set up the fault-free dataset for sensors.

In Step 320, the computing device 200 generates a faulty dataset by introducing instantaneous sensor signal errors e _i into the fault-free data, and then keeps this faulty dataset in the memory block 221 b. In one example, three types of instantaneous signal errors e _i are taken into consideration, that is, bias, drift, and precision degradation. Bias (B) refers to a positive or a negative constant difference throughout a predetermined period for a parameter/asensor. Drift (D) refers to a difference the magnitude of which increases or decreases with time. Precision degradation (P) is determined from a Gaussian distribution function with a mean value zero and a standard deviation. By introducing these instantaneous sensor signal errors, a faulty dataset is set up.

In Step 330, the computing device 200 runs a preset system model with the fault-free dataset and the faulty dataset as inputs, to get a modeled dataset. The system model may be previously built up and stored in the memory block 222. In one example, the system model is set up through deep learning (e.g. training an artificial neural network model) using time-independent operating parameters and the time-dependent sensor data, the latter of which is the fault-free dataset. The setup system model may accurately characterize the performance of a target chiller plant from which the training data are collected, without knowing the detailed design of the chiller plant.

In Step 340, the computing device 200 further build up a fault pattern database by applying a data-mining method to the modeled dataset, then the fault pattern database including both fault-free profiles and faulty profiles. The “profile” here refers to one or more feature measurements characterizing fault-free cases and faulty cases. The “profile” may be different from one data-mining method to another.

In Step 350, the computing device 200 evaluates and reconstructs those sensor data for evaluation when a faulty profile fits well with the profile of the sensor data for evaluation.

With the SFM method 300, the computing device 200 may identify faulty sensor signal readings and reconstruct these sensor signal readings with reference to faulty cases. By training an ANN model with fault-free dataset to build up a system model for a chiller plant, there is no need to fully understand the specific design and control strategy applied which may differ significantly from a chiller plant to another chiller plant in actual systems. The reconstructed sensor signal readings may be used by the computing device or BMS to fulfill its operation control strategy of the chiller plant, or optimize its maintenance strategy, or even for further analysis. This kind of fault reconstruction is also valuable, as some fault sensors (such as the water flow sensors) are not easy to be replaced in actual installations due to their severe disruption to the normal system operation. With the SFM method 300, a faulty sensor reading may be reconstructed or corrected without replacing the fault sensors immediately, which provides BMS with more reliable inputs and thus ensuring a more robust chiller plant from control and management aspect.

The SFM method 300 is based on deep learning and pattern recognition, which is in theory applicable to all sensors of each chiller plant, of course requiring a considerable amount of computing resource. With this observation, the inventor of the present disclosure further figures out that it would be more efficient and effective if the SFM method 300 could only apply to those sensors to which the performance of the chiller plant is more sensitive than others. These sensors are called as critical sensors (CS) in the present disclosure.

The critical sensors of a chiller plant may be identified by various methods. One easy way is to identify those critical sensors based on the experience or knowledge of chiller plants. For example, for a water-cooled system, the sensor for detecting temperature of chilled water at return header and the sensor for detecting the temperature of cooling water at return/supply header are taken as the critical sensors upon knowledge or experience. The non-critical sensors, on the chiller water side, may for example include those sensors installed at the chilled water inlet and outlet of each chiller as well as that at the chilled water supply header. On the cooling water side which only applied to water-cooled systems, the non-critical sensors may for example include those sensors placed at the cooling water inlet and outlet of each chiller as well as the water inlet and outlet of each cooling tower. An alternative way to identify critical sensors is to conduct Sensitivity Analysis, one example of which is shown in FIG 4 and investigates the chiller plant’s sensitivity to one parameter/sensor signal reading by identifying the difference of the performance of the chiller plant over the entire parameter/sensor range. The detailed description with reference to FIG 4 is in later part of the present disclosure.

FIG 4-7 illustrate the detailed examples of each step of SFM method 300, where the deep learning and pattern recognition are only used to evaluate and reconstruct those critical sensors under fault situation.

FIG 4 shows a flow chat of a pre-handling method, which may be done before the Step S310 and the output of which may be taken as inputs to method 300. The pre-handling method is to build up a system model to simulate a specific chiller plant and identify those critical sensors that are suitable for being evaluated or reconstructed by SFM 300.

The flow chart of FIG 4 starts from Step 410 to receive or read out plant data that are logged in a BMS system at different sites. In this example, a water-cooled system for chiller plant is considered. Here, the plant data refers to those which describe the plant configuration and characteristic that do not vary with time as well as others which show the performance of the plant at different time. Hence, two separate data files are required, namely a time-independent input file and a time-dependent input file.

For simplicity, the time-independent inputs are also referred to as parameters of a chiller plant that do not vary with time and can be initially obtained by manual input or reading from a configuration/schedule file of the chiller plant. For example, the parameters for a chiller plant may include the number of its chillers/chilled water pumps/cooling towers/cooling water pumps (N _ch, N _chwp, N _ct, N _cwp) , design chilled water supply temperature (T _chws, d) , design chilled water flow rate (m _chw, d) , design cooling capacity (Q _e, d) , maximum cooling water supply temperature (T _cws, max) and so on. The time-dependent inputs refer to those sensor signal readings vary with time. For example, the time-dependent input may include temperature of chilled water at its supply header (T _chws) , temperature of chilled water at its return header (T _chwr) , temperature of cooling water at its supply header (T _cws) , temperature of cooling water at its return header (T _cwr) , mass flow rate of chilled water (m _chw) , mass flow rate of cooling water (m _cw) , or the like upon specific situation or requirements.

Data Refinement

In Step S420, the plant data, i.e. the logged data from BMS system, are being refined to ensure their reliability. The data refinement step S420 may be preferably grouped into any of the four categories, namely,

- conversion of non-numeric data to numeric (R0) ,

- check for missing data (R1) , e.g. by linear interpolation,

- check for data out-of-range (R2) , e.g. by reset the out-of-range data at the upper/lower limit of its range, and

- check for data conflict with equipment on/off status (R3) , e.g. by following control logic or physical principles.

Upon completion of the data refinement in Step S420, various plant data files are created to record the data refinement checking results in memory 220. The refined data may be recorded in memory block 221.

With the data refinement, the plant data to be used for system modeling or further analysis may follow a standard data format or the same data structure, even if the plant data are collected from a BMS system at different sites. With the data refinement, the incomplete data, the conflict data or the data out of a reasonable range may be identified at an early stage, and thus the refined plant data are more reliable. This would be beneficial to forming an accurate system model from the starting point.

Among the refined plant data, the time-dependent input data is divided into two parts. The first part is taken as fault-free sensor data for system modeling and fault-pattern database building. For example, the time-dependent input data collected from sensors within a predetermined period from their installation are taken as fault free. In an example, the sensor signal readings within the first 6 months of sensor operation is taken as the fault-free data, which would form a fault-free dataset as the input of Step 310 of method 300. The fault-free dataset may be kept in memory sub-block 221a. The latter part, e.g. the time-dependent input data collected from sensors later than the first 6 months, is taken as the actual sensor signal readings, which may be stored in memory sub-block 221c for evaluation and reconstruction in Step S350 of method 300.

System Modeling

To simplify the computing complexity, in this example, deep learning and pattern recognition algorithms are only applied to those critical sensors, then less information is needed.

In Step S430, a system model is built up by training an ANN model with the parameters of a chiller plant and the fault-free data obtained after data refinement. It is important for an ANN model to define parameters at its input and output layers and its structure. In one example, the inventor involves previous sensor signal readings in the input player of the ANN model.

Thus, the training data of the ANN model at input layer includes not only the sensor signal readings at the current time step, but also the sensor signal readings in previous time steps. In this way, the impact of operating situation in previous time steps on the chiller plant performance in current time step is also taken into consideration. Preferably, the training data of the ANN model at the output layer mainly highlights the key operating parameters and sensor readings of those major equipment and the system. Table 1 shows an example of training dataset at the input and output layers of ANN model, where three supply and return header streams are considered.

Table 1 Summarized input and output layer of the system model

Preferably, all input and output data for training the above ANN model may be normalized between the minimum and maximum values of their parameters or sensor signal readings in the time-dependent input database. For any input/output x, the respective normalized value is given by X.

The structure of the ANN model can substantially affect the accuracy of the model. In the example of FIG 4, an ANN with two intermediate layers is employed, the first one having more than 400 neurons, e.g. around 460 neurons, and the second one having more than 300 neuros, e.g. around 360 neurons. Upon various requirements, the structure of ANN model may be different, not limited to the above. In general, the use of more intermediate layers or more neurons in each layer can improve the precision of the model but at the expense of longer computation time. Those skilled in the art may achieve other appropriate structure of ANN model upon specific requirements.

To train the ANN model, 70%of the time steps of the fault-free data (in the time-dependent input file) is randomly selected in one example. The remaining is used for model validation (e.g. 15%) and testing (e.g. 15%) . For a building which is not operating in 24 hours basis, the performance of the chiller system during the night period usually appears to be intermittent which influences the effectiveness of the system modeling. Consequently, in such case, the operating plant data between 10: 00 pm and 8: 00 am is taken out, not being used in SFM.

By training the ANN model with the selected input and output data, a system model is built up in Step S430 and stored in memory sub-block 222.

Sensor Selection

Preferably, in the example of FIG 4, the computing device 200 may further identify the critical sensors by conducting sensitivity analysis in the optional Step S440. Alternatively, the sensitivity analysis in the optional Step S440 may be used to double check if the manually determined critical sensors really affect the system performance to a greater extent. There are various ways to conduct the sensitivity analysis. In this example, the inventor uses the system model setup in Step S430 to analyze the sensitivity of the system performance over the entire range of each sensor signal reading.

Specifically, for each selected normalized sensor signal reading X, a base value (X _base) is defined which is taken as a median value in the input dataset. Then, the respective maximum (X _max) and minimum (X _min) values are found within the input dataset to identify the entire range of the sensor signal reading. When investigating the chiller plant sensitivity to one sensor signal reading, base values of the other parameters and other sensors are adopted. The differences in the system performance, which may be indicated by a system coefficient of performance COP _sys, over the entire range of the sensor signal reading are compared. A greater difference in COP _sys indicates a higher sensitivity of the chiller plant performance to that sensor, and thus the sensors to which the chiller plant is more sensitive may be determined as critical sensors. The manual input about critical sensors may be replaced by the sensitivity analysis in Step S440. Here, the COP _sys may take both the input and output of the built-up system model into consideration, thus reflecting the performance of the modeled chiller plant precisely.

Faulty dataset building

Now refer back to FIG 3. In a preferable example, a faulty dataset is built up with reference to the fault-free dataset in Step S320 as follows. In the example, faulty dataset is built up by introducing instantaneous signal error e _j into the fault-free data x _ij, ff, that is, faulty data x _ij, ft=x _ij, ff+e _ij. For example, three types of instantaneous signal error e _i may be considered, namely:

- instantaneous signal error for bias (B) , which is a positive or negative constant throughout a whole evaluation period;

- instantaneous signal error for drift (D) , the magnitude of which increases or decreases with time;

- instantaneous signal error for precision degradation (P) , which is determined from a Gaussian distribution function with mean value zero and standard deviation.

To build up the faulty dataset, different fault cases were considered. This includes all the three fault types with different fault strengths. In one example, two types of sensors are involved, namely temperature and flow sensors. Table 2 summarizes the variation of fault cases investigated in the example. For each sensor signal reading, there would be totally 38 fault cases. Multiple sensor faults can also be dealt with at the same time, while the number of fault combinations would be increased (totally 1, 520 faulty cases for two sensors only) . The obtained faulty dataset is recorded in memory sub-block 221b. The SMF of present disclosure is applicable to considering multiple sensor faults.

Table 2 Definition of different fault cases for both temperature and flow sensors

1 Maximum system flow.

Fault pattern database building

In Step S330, the system model set up in Step 430 is run by taking the fault-free dataset and the faulty dataset as inputs respectively to obtain a modeled dataset. The modeled dataset reflects the performance of the chiller plant under the fault-free situations and faulty situations.

In Step S340, a data-mining method is further applied to the modeled dataset to build up a fault pattern database including fault-free profiles and faulty profiles, where the “profile” refers to one or more feature measurements characterizing fault-free cases and faulty cases.

In one example, the data-mining method is a clustering method, e.g. k-means clustering, and the fault pattern database may be for example represented by a total centroid score CS _t, at any time step i, which is given by

where

X _i(i = 1, 2, …t _max) is a normalized modeled data, t _max is the maximum time step in a predetermined period for the data-mining,

c _k(k = 1, 2, …N) is a centroid of the kth cluster, N is the number of clusters and given by a rule-of-thumb;

G _k(k = 1, 2, …N) is the kth clustered subset;

n _k is the number of time steps in the clustered subset G _k containing that at time step i;

‖ ‖is Euclidean distance.

For example, if the model training data is on a weekly basis, respective fault-free and faulty profiles are generated to characterize the weekly performance of the chiller plant. With Ns critical sensors (preferably selected or double checked in Step 440) , a normalized modeled dataset for fault free cases X _ff= {X _i, j, ff, i=1, 2, .., t _max, j=1, 2, .., Ns} is built, where t _max is the maximum time step in a week, and ff refers to the fault-free cases. Then, k-means clustering is applied to group the fault-free modeled data into a number of clusters N, where N is given by a rule-of-thumb. With a centroid dataset of the clusters C= {c _k, j, k=1, 2, .., N, j=1, 2, .., Ns} computed, the clustered dataset for fault-free cases G _ff= {G _k, ff, k=1, 2, .., N } is determined where U {G _k, ff} = X _ff. In this case, the fault-free total centroid score at any time step (CS _t, i, ff) is then given by

Then, preferably, all the CS _t, i, ffis divided by the weekly mean fault-free total centroid score to get a normalized the CS _t, i, ff. Thus, the normalized CS _t, i, ff represents the fault-free profiles.

In the similar way, the faulty profiles may be obtained by clustering and represented by

where, ft refers to the faulty cases. Same as that for faulty-free profiles, all the CS _t, i, ft may be divided by the weekly mean faulty total centroid score to get the normalized CS _t, i, ft.

In step S340, other data-mining method may be used to partition the modeled data for fault-free cases and those for faulty cases. For example, fuzzy c-means clustering, QT clustering or locality-sensitive hashing may be used. Meanwhile, the profile may be represented by another feature measurement other than the centroid score mentioned here.

By applying data-mining algorithm to modeled data, a fault pattern database including both fault-free profiles and faulty profiles is set up and may be kept in the memory 223. The fault pattern database may be taken as a good reference database for evaluation.

Evaluation and reconstruction for critical sensors

FIG 5 shows a detailed example to fulfill the Step S350 of SFM method 300. That is, a detailed example to evaluate those sensor signal readings of critical sensors and reconstruct them in case of faulty is provided.

The method of FIG 5 starts from Step S510, in which the sensor dataset for evaluation is received. The sensor dataset for evaluation may be, for example, the latter part of refined time-dependent data stored in memory sub-block 221c. For example, the sensor dataset for evaluation may be the sensor signal readings collected from those critical sensors after the first 6 months of operation. The sensor dataset for evaluation may include various signal readings of critical sensors over an evaluation period. For example, the evaluation period may be around one month, and then the sensor dataset for evaluation covers the sensor signal readings collected in the whole month from those critical sensors, e.g. at the both input and output layers of the system model.

For example, if respective fault-free and faulty profiles are generated as mentioned above on a weekly base, the weekly evaluation profile should be generated in the same way. That is, the evaluation profile may be presented by the centroid score as well. In Step S520, the centroid score of evaluation profile is the CS _t, i, e, which is given by

where, X _e= {X _i, j, e, i=1, 2, .., t _max, j=1, 2, .., Ns} , X _i, j, e refers to sensor signal reading of j _th critical sensor at the time step i for evaluation. Same as that for faulty-free profiles, all the CS _t, i, e may be divided by the weekly mean evaluation total centroid score to get a normalized CS _t, i, e. The evaluation profiles for critical sensors may be recorded in the memory sub-block 224.

In Step 530, the evaluation profile for each evaluation week would be compared with all the weekly fault-free and faulty profiles week-by-week, until all evaluation profiles over the entire evaluation period are compared. The fitness of the profiles is determined, for example, by using the Euclidean distance between the fault-free/faulty profiles CS _t, i, ff/CS _t, i, ft and the evaluation profile CS _t, i, e. For example, a fault profile with the smallest Euclidean distance to the evaluation profile would be identified for each evaluation week. The faulty case corresponding to the fault profile which is of the highest frequency of occurrence throughout the whole evaluation period would be taken as the fault diagnosis result for the selected critical sensor for evaluation. The faulty case would include the fault type and the fault value. The step 530 may be used to evaluate multiple sensors at the same time.

With the fault evaluation of all the critical sensors completed, a reconstruction file would be generated which records for example the starting time (Year, month and day) , fault type ("B" for bias, "D" for drift and "P" for precision degradation) and value of the fault reconstructions. For precision degradation, the reconstruction value may be set to zero. For bias and drift, it may be simply the negative of the fault value. Table 3 shows an example of the reconstruction results of critical sensors. The reconstruction results would be further used in the fault evaluation of non-critical sensors to reconstruct the respective non-critical sensor signals when appropriate. The reconstructed sensor data may be stored in sub-block 225 of the memory.

Table 3: An example of the reconstruction results of critical sensors.

Fault evaluation for non-critical sensors

The methods illustrated in FIG 3-5 employing system modeling, data-mining and pattern recognition, which is very suitable for critical sensors to which the chiller plant is more sensitive. To lower the complexity, considering the chiller plant performance is less sensitive to those non-critical sensors, the inventor further proposes to take bias fault into account only for those non-critical sensors, e.g. temperature bias of the non-critical sensors may be computed by using energy and mass balance method, as shown in FIG 6.

The method of FIG 6 starts from Step S610, in which all the refined time-dependent data stored in memory block 211 are received, including the fault-free dataset and the evaluation dataset for critical sensors as well as the time-dependent data for non-critical sensors.

In Step 620, the reconstructed signal readings of critical sensors are received and taken as BASE values, or reference values. In other words, it is assumed that the reconstructed signal reading of a critical sensor is correct and could be taken as a starting/reference point for the evaluation of those non-critical sensors which are logically related to the critical sensors.

In Step 630, with reference to the BASE values, the corresponding temperature bias at various locations are calculated or determined. For example, on the chilled water side, the temperature bias at different locations is determined by using the chilled water return temperature as the BASE value, which was reconstructed in Step S530. For the cooling water side (only applicable to water-cooled system) , the cooling water supply and return temperatures are used as the BASE values for calculating the respective temperature bias at various locations.

For each selected time step, the computation of the bias is only made to those sets of equipment which are in operation. Respective trend data of the temperature bias for the non-critical sensors are then generated in Step S640.

In Step S650, it is to smooth or flatten the trend data in order to determine the representative bias throughout the whole evaluation period. This may be accomplished in three stages. In the first stage, the trend profiles are segmented in various groups by using the moving average approach. A moving average was computed along each trend profile until the instant when the trend data value deviated from the moving average by a certain value like 0.3 ℃. A new group was then formed with the new moving average calculated until the end of the trend profile. In the next stage, the trend profile of each group was leveled based on the group average value. The final step involved the merging of adjacent groups with group average values differed by less than another certain value like 0.2 ℃, starting from the first instant of the trend profile. The new average values for the merged groups would then be re- calculated. The process repeated until there was no more group merging and the last bias value is taken as the final output.

Upon completion of the calculation, two data files could be created. The first file may summarize the checking results of all the sensors including both critical and non-critical ones. For critical sensors, it might be bias, drift or precision degradation while it is always bias only for the non-critical sensors. For air-cooled systems or in case no cooling water flow information is available, only those parameters involving chilled water would be shown. If the usable dataset is limited or when no chilled water flow signal was detected, non-critical sensor checking would not be conducted. Only the results for critical sensors would be indicated, and no trend plot would be generated. The first file may look similar to Table 3. The second file may indicate the temperature bias trend plots (both original and flattened, as shown in FIG 7) of all the non-critical sensors that could be calculated. The final reconstruction report may be kept in memory sub-block 225.

A novel SFM and sensor data reconstruction solution for chiller plants through big data analytics is provided with reference to the above examples. The most preferable solution is devised which employ deep learning algorithm to model the chiller plant performance and clustering method to predict various types of sensor faults for critical sensors; while thermodynamics approach to determine the bias for non-critical sensors. A full summary report for fault reconstruction of both critical and non-critical sensors, as well as trend plots for non-critical sensors, could be generated for the maintenance staff to take any remedial actions. Various chiller plant logged data as well as site checking reports are employed to validate the solution. The proposed SFM solution may be considered as not only a robust choice, but also a foundation for future because of its higher accuracy and reliability and less manpower.

It should be understood that the foregoing relates only to certain examples of the disclosure and that numerous changes can be made therein without departing from the scope of the disclosure as defined by the following claims. It should also be understood that the disclosure is not restricted to the illustrated examples and that various modifications can be made within the scope of the claims.

Reference numerals

100 Chiller Plant

110 Chilled water circuit

112 Chillers

116 Chilled water pumps

120 Cooling water circuit

124 Cooling towers

126 Chilled water pumps

140 Sensors

20 Sensor management system

200 Computing device

210 Processor

220 Memory

221～225 Memory sub-block

280 BMS

Claims

A method of sensor fault management for a chiller plant, comprising:

obtaining (S310) a fault-free dataset of sensors for the chiller plant;

generating (S320) a faulty dataset by introducing instantaneous sensor signal errors into the fault-free dataset;

running (S330) a system model with the fault-free dataset and the faulty dataset as inputs to obtain a modeled dataset, wherein the system model is set up by deep learning with operating parameters and the fault-free dataset to characterize performance of the chiller plant;

building up (S340) a fault pattern database including fault-free profiles and faulty profiles based on the modeled dataset, where profile of the fault-free profiles or the faulty profiles refers to one or more feature measurements characterizing fault-free cases or faulty cases;

evaluating and reconstructing (S350) sensor data for evaluation, wherein the sensor data for evaluation is identified as fault and reconstructed when a faulty profile fits with the sensor data.
The method according to claim 1, wherein,

the fault-free dataset only includes data collected from those critical sensors, to which the performance of the chiller plant is more sensitive than other sensors under fault situation.
The method according to claim 2, wherein:

a sensor is identified as the critical sensor if a change of the performance of the chiller plant over an entire data range of the sensor is greater than a threshold, while other sensors are being kept at their median value.
The method according to claim 1, wherein the operating parameters and fault-free data for deep learning comprise: the fault-free data of sensors in current time step and those fault-free data of sensors in previous time step;

preferably, the operating parameters and fault-free data at input layer of the deep learning are any of or any combination of followings: outdoor temperature, temperature of chilled water at return header, temperature of chilled water at supply header, temperature of cooling water at return header, temperature of cooling water at supply header, mass flow rate of chilled water, mass flow rate of cooling water, or mass flow rate of cooling tower air;

preferably, the operating parameters and fault-free data at output layer of the system model are any of or any combination of followings: mass flow rate of chilled water, mass flow rate of cooling water, mass flow rate of cooling tower air, power consumption of chilled water pump, power consumption of cooling tower, power consumption of cooling power pump, number of chillers, number of chilled water pumps, numbers of cooling towers, number of cooling power pumps, temperature of chilled water at return header, or temperature of chilled water at supply header.
The method according to claim 1, further comprising the step of:

Refining data collected from sensors to form the fault-free dataset by implementing any of the steps of:

converting non-numeric data to numeric data;

checking for missing data,

checking for data out-of-range, or

checking for data conflict with equipment on/off status.
The method according to claim 1, wherein the instantaneous signal errors comprise any of followings:

- instantaneous signal error for bias, which is a positive or negative constant throughout a whole evaluation period;

- instantaneous signal error for drift, the magnitude of which increases or decreases with time;

- instantaneous signal error for precision degradation, which is determined from a Gaussian distribution function with mean value zero and standard deviation.
The method according to claim 1, wherein the step of evaluating and reconstructing sensor data for evaluation comprises:

obtaining a plurality of evaluation profiles over an evaluation period, each evaluation profile is based on signal readings of one or more sensors selected for evaluation at a predetermined evaluation time interval, the predetermined evaluation time interval is same as the time interval for fault pattern database;

comparing the evaluation profile for each evaluation time interval with all of the faulty-free profiles and faulty profiles of the fault pattern database one by one;

for each evaluation interval, identifying a fault profile that fits well with the evaluation profile,

taking a fault case with highest frequency of occurrence throughout the whole evaluation period as a fault diagnosis result for the selected sensors.
The method according to claim 1, wherein the data-mining method is a clustering algorithm to group the modeled data into clusters and the profile comprises a feature measurement describing characteristic of each cluster.
The method according to claim 8, wherein the clustering algorithm is k-means clustering algorithm and the profile comprises a total centroid score at any time step i , CS _t, i, given by

where

X _i (i = 1, 2, …t _max) is a normalized sensor signal, t _max is the maximum time step in a predetermined period for the data-mining,

c _k (k = 1, 2, …N) is a centroid of the kth cluster for the fault-free cases, N is the number of clusters and given by a rule-of-thumb;

G _k, ff (k = 1, 2, …N) is the kth clustered data subsets for the fault-free cases;

n _k is the number of time steps in the kth clustered subset G _k, ff containing that at time step i.
The method according to claim 9, wherein a fault profile fits well with the sensor data when having a smallest Euclidean distance between the faulty profile CS _t, i, ft, and the evaluation profile of the sensor data CS _t, i, e.
The method according to claim 2, further comprises: ,

With the reconstructed sensor data, calculating bias for non-critical sensor at different locations of the chiller plant,

generating trend data of the bias for the non-critical sensors;

determining a representative bias throughout a whole record period for the non-critical sensors by flattening the trend data.
The method according to claim 11, the step of flattening the trend data comprises:

segmenting the trend data into various groups by using a moving average approach;

leveling the trend data of each group based on a group average value;

merging of adjacent groups with group average values differed by less than a predetermined threshold;

recalculating a new average value for each merged group and repeating the merging until no more group merging.
A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of any of the methods according to claims 1 to 12.
A computer-readable medium having stored thereon the computer program of claim 13.
A system for sensor fault management for a chiller plant, comprising:

a computing device (200) , configured to communicate with a Building Management System that controls a chiller plant installed with a plurality of sensors, wherein the computing device is configured to,

receive a fault-free dataset of the sensors;

generate a faulty dataset by introducing instantaneous sensor signal errors into the fault-free dataset;

run a system model with the fault-free dataset and the faulty dataset as inputs respectively to obtain a modeled dataset, wherein the system model is set up by deep learning with operating parameters and the fault-free data to characterize the performance of the chiller plant;

build up a fault pattern database including fault-free profiles and faulty profiles by applying a data-mining method to the modeled dataset, where the profile refers to one or more feature measurements characterizing fault-free cases or faulty cases;

evaluate and reconstruct sensor data for evaluation, wherein the sensor data for evaluation is identified as fault and reconstructed when a faulty profile fits with the sensor data.
The system according to claim 15, wherein,

the fault-free dataset only includes data collected from those critical sensors, to which the performance of the chiller plant is more sensitive than other sensors under fault situation;

preferably, a sensor is identified as the critical sensor if there is a greater change of the performance of the chiller plant over an entire data range of the sensor while other sensors are being kept at their median value.
The system according to claim 15, wherein the operating parameters and fault-free data for deep learning comprise: the fault-free data of sensors in current time step and those fault-free data of sensors in previous time step;

preferably, the operating parameters and fault-free data at input layer of the system model are any of or any combination of followings: outdoor temperature, temperature of chilled water at return header, temperature of chilled water at supply header, temperature of cooling water at return header, temperature of cooling water at supply header, mass flow rate of chilled water, mass flow rate of cooling water, or mass flow rate of cooling tower air;

preferably, the operating parameters and fault-free data at output layer of the system model are any of or any combination of followings: mass flow rate of chilled water, mass flow rate of cooling water, mass flow rate of cooling tower air, power consumption of chilled water pump, power consumption of cooling tower, power consumption of cooling power pump, number of chillers, number of chilled water pumps, numbers of cooling towers, number of cooling power pumps, temperature of chilled water at return header, or temperature of chilled water at supply header.
The system according to claim 15, wherein the computing device (200) is further configured to:

obtain a plurality of evaluation profiles over an evaluation period, each evaluation profile is based on signal readings of a sensor selected for evaluation at a predetermined evaluation time interval, the predetermined evaluation time interval is same as the time interval for fault pattern database;

compare the evaluation profile for each evaluation time interval with all of the faulty-free profiles and faulty profiles of the fault pattern database one by one;

for each evaluation interval, identify a fault profile that fits well with the evaluation profile,

take a fault case with highest frequency of occurrence throughout the whole evaluation period as a fault diagnosis result for the selected sensor.
The system according to claim 15, wherein the data-mining method is k-means clustering algorithm and the profile comprises a total centroid score at any time step i , CS _t, i, given by

where

X _i (i = 1, 2, …t _max) is a normalized sensor signal, t _max is the maximum time step in a predetermined period for the data-mining,

c _k (k = 1, 2, …N) is a centroid of the kth cluster for the fault-free cases, N is the number of clusters and given by a rule-of-thumb;

G _k, ff (k = 1, 2, …N) is the kth clustered data subsets for the fault-free cases;

n _k is the number of time steps in the kth clustered subset G _k, ff containing that at time step i.
The system according to claim 16, wherein the computing device (200) is further configured to:

with the reconstructed sensor data as a base value, calculating bias for non-critical sensor at different locations of the chiller plant,

generate trend data of the bias for those non-critical sensors;

determine a representative bias throughout a whole record period for the non-critical sensors by flattening the trend data.