A fault detection and diagnosis method for air handling units with different node granularities

Abstract

The present invention discloses a method for detecting and diagnosing faults of an air handling unit with different node granularities, comprising the following steps: (1) calculating the fault level and the minimum attribute reduction symptom set of the fault set of the air handling unit based on the knowledge granularity; (2) determining the symptom nodes and fault nodes in the Bayesian network according to the calculation results of the fault level sorting and the minimum attribute reduction symptom set; (3) constructing a Bayesian network structure through a scoring search algorithm according to the symptom nodes and fault nodes of the Bayesian network, and obtaining the structure of the Bayesian network with different fault granularities; (4) performing data-driven parameter learning on the constructed Bayesian network structure; (5) using the Bayesian network structure after parameter learning to detect and diagnose faults of the air handling unit. Compared with the traditional method, the present invention has higher versatility and more accurate diagnosis results.

Description

Technical Field

[0001] The invention belongs to the field of fault detection and diagnosis, and in particular relates to a method for fault detection and diagnosis of an air handling unit with different node granularities. Background Art

[0002] Air Handling Unit (AHU) is an important part of the HVAC system, which can adjust the air temperature and humidity in each room and area. AHU failure will cause parameters such as air supply temperature and air supply flow to deviate from the set values, thus affecting indoor comfort. A long-term faulty operation will also shorten the service life of the equipment and lead to a large amount of energy waste. Manually repairing many AHU faults is not only time-consuming and laborious, but also requires high professionalism from technicians. Therefore, the research on automatic fault detection and diagnosis methods has important scientific value and engineering significance.

[0003] A Chinese patent document with publication number CN106871389A discloses an active fault diagnosis method for a variable air volume air handling unit, which uses expert rules to detect fault symptoms with multiple fault causes, and then causes the variable air volume air handling unit to enter a pre-set active fault diagnosis step, eliminate potential fault causes item by item, and separate and diagnose the variable air volume air handling unit fault.

[0004] A Chinese patent document with publication number CN109813542A discloses a fault diagnosis method for an air handling unit based on a generative adversarial network. The fault sample generation model is trained by sampling real fault samples to learn the distribution of real fault samples. The fault diagnosis model is trained using the generated samples to detect the data generated during the operation of the air handling unit in reality to diagnose whether the equipment has a fault.

[0005] However, the active fault diagnosis method relies on domain experts, and the structure and parameters of the model have clear physical meanings. However, the AHU system is large in scale and has a wide variety of equipment. It takes a lot of manpower and time to check each device one by one. At the same time, it is limited by the professional quality of the operation and maintenance personnel, and the diagnostic accuracy is low. The fault diagnosis method based on the generative adversarial network can automatically learn and train, and has high diagnostic accuracy within the data set range, but it relies on a large amount of fault label data and has high requirements on data quality. It is uninterpretable and has poor extension, making it difficult to transplant to other systems. HVAC systems are extremely diverse and complex. The above two patented methods are developed for equipment / systems under specific working conditions and are difficult to meet application requirements.

[0006] Bayesian networks have flexible network structures and probabilistic parameters. They can imitate expert diagnostic thinking and perform probabilistic reasoning when information is incomplete and uncertain. Due to their strong interpretability, uncertainty modeling, and convenient integration of multi-source information, they are increasingly widely used in the field of fault diagnosis. However, due to the poor versatility of traditional Bayesian network construction methods, they are greatly limited in practical applications. Taking air handling unit fault detection and diagnosis as an example, when the sensors arranged in the air handling unit fail or migrate to a building with the same equipment configuration but a smaller number of sensors, the sensor arrangement changes, the collected symptom information also changes, and some symptom nodes in the Bayesian network are deleted. As a result, some faults with similar symptoms cannot be distinguished due to the lack of key diagnostic information, resulting in misdiagnosis. At the same time, as the collected symptom information decreases, the granularity of the faults that can be diagnosed gradually becomes coarser, and even cannot be diagnosed. Generally, we call the degree of detail of a fault node granularity. Faults with finer granularity correspond to higher levels and can provide more detailed diagnostic results, which are called fine faults. Conversely, faults with lower levels correspond to coarser granularity and provide broader diagnostic results, which are called coarse faults. Usually, coarse faults and their corresponding fine faults have a physical representation relationship.

[0007] Therefore, it is very important to develop a fault detection and diagnosis method that can dynamically adjust the faulty nodes according to the richness of different symptom information. However, there is still a lack of research on Bayesian network fault detection and diagnosis methods that can automatically select different nodes. Summary of the invention

[0008] The present invention provides an air handling unit fault detection and diagnosis method with different node granularities, which has higher versatility and more accurate diagnosis results.

[0009] A method for detecting and diagnosing faults of an air handling unit with different node granularities comprises the following steps:

[0010] (1) Based on knowledge granularity, the fault level and minimum attribute reduction symptom set of the fault set of the air handling unit are calculated;

[0011] (2) Determine the symptom nodes and fault nodes in the Bayesian network based on the calculation results of the fault level sorting and the minimum attribute reduction symptom set;

[0012] (3) Based on the symptom nodes and fault nodes of the Bayesian network, a Bayesian network structure is constructed through a scoring search algorithm to obtain a Bayesian network structure with different fault granularities;

[0013] (4) Perform data-driven parameter learning on the constructed Bayesian network structure;

[0014] (5) The Bayesian network structure after parameter learning is used to detect and diagnose air handling unit faults. The abnormal symptoms of the air handling unit are input and the posterior probability of each fault occurring is output.

[0015] The present invention proposes a method for detecting and diagnosing air handling unit faults with different node granularities, which can adaptively divide the fault level and the relationship between the coarse and fine granularities, and automatically compare the existing symptom information with the necessary diagnostic information, and finally retain the appropriate fault nodes to ensure the accuracy and precision of the fault diagnosis results. Here, accuracy means that the result of the Bayesian network fault diagnosis is correct, and precision means that on the basis of ensuring accuracy, the faults diagnosed by the Bayesian network are as high-level and fine-grained as possible.

[0016] The specific process of step (1) is as follows:

[0017] (1-1) Obtain a hierarchical set of historical data, including all fault information and symptom information of the air handling unit;

[0018] (1-2) Preprocessing the obtained hierarchical set of historical data;

[0019] (1-3) Randomly select a fault as the decision attribute, take other faults as conditional attributes, perform minimum attribute simplification on the selected fault, and obtain the minimum attribute simplification fault set, knowledge granularity, and relative knowledge granularity;

[0020] (1-4) When the relative knowledge granularity of the selected fault is 0, it means that the selected fault and the minimum attribute reduction fault set have a hierarchical relationship. If the knowledge granularity of the selected fault is smaller than the knowledge granularity of the minimum attribute reduction fault set, then the fault is a coarse fault of the minimum attribute reduction fault set, otherwise, it is a fine fault.

[0021] (1-5) For all faults, loop through steps (1-3) and (1-4) to obtain the hierarchical relationship of all faults based on the calculation results;

[0022] (1-6) Randomly select a fault as the decision attribute, use all the symptom data in the historical data layer set as the conditional attributes, perform minimum attribute simplification on the selected fault, and obtain the minimum attribute simplified symptom set;

[0023] (1-7) For all faults, loop step (1-6) to obtain the minimum attribute reduction symptom set for each fault.

[0024] The specific process of minimum attribute reduction is:

[0025] Step 1: Determine the decision table S =<U,C∪D,V,f> , where U is the domain, C and D are the condition attribute set and decision attribute set respectively, V is the range, f is an information function, and an empty reduction set RED is defined at the same time U ;

[0026] Step 2: Select any attribute a from the conditional attribute set C j , calculate the property a j On the importance of internal attributes of condition attribute set C relative to decision attribute set D

[0027]

[0028] if Then a j Add to RED U In the above example, we traverse all the attributes in C and obtain the reduced set RED that is initially screened. U_B ;

[0029] Step 3, if the reduction set RED U_B It has different partitioning capabilities from the total condition attribute set C, that is, GP U (D|RED U_B )≠GP U (D|C); Subtract the reduction set RED from the condition attribute set C U_B Get the attribute set C-RED U_B , from the attribute set C-RED U_B Pick any attribute a j , calculate the property a j About the conditional attribute set RED U_B The importance of external attributes relative to the decision attribute set D

[0030]

[0031] Traversing the attribute set C-RED U_B Among all the attributes in , select the attribute a0 with the greatest external attribute importance and add it to RED U_B middle;

[0032] Step 4: Repeat step 3 until the reduced set RED is obtained. U_B It has the same partitioning capability as the set of all conditional attributes C;

[0033] Step 5. From the new attribute set RED U_B Any attribute a j Get the attribute set RED U_B -{a j}, if the attribute set RED U_B-{a j} has the same partitioning ability as the entire condition attribute set C, then a j From the attribute set RED U_B Delete from a and loop through a j , and obtain the final minimum attribute reduction set.

[0034] In steps (1-3), the calculation formulas for knowledge granularity and relative knowledge granularity are:

[0035]

[0036] GP U (D|B)=GP U (B)-GP U (B∪D)

[0037] In the formula, GP U (B) is the knowledge granularity GP of set B U (B), GP U (D|B) is the relative knowledge granularity of set D relative to set B. Given a decision table S = (U, A = C∪D, V, f), U is the domain, if U / B={X1,X2,...,X m}, there are m equivalence classes, |X i | represents the number of elements in the i-th equivalence class, where i = 1, 2…m.

[0038] The specific process of step (2) is as follows:

[0039] (2-1) According to the arrangement of sensors in the air handling unit to be diagnosed, historical data are collected to build a set, the data set is preprocessed, and the known signs and sign nodes of the Bayesian network are determined;

[0040] (2-2) Randomly select a fault and determine whether to select the node in the Bayesian network based on the relationship between the minimum attribute reduction symptom set of the fault and the known symptom set: if the minimum attribute reduction symptom set of the fault is a subset of the known symptom set, then introduce the fault into the Bayesian network fault node set, otherwise do not introduce it;

[0041] (2-3) Loop through step (2-2), traverse all faults, and obtain the set of fault nodes constructed by the Bayesian network;

[0042] (2-4) Based on the fault hierarchy sorting, overlapping faults are screened out: if the introduced faults contain both coarse faults and all fine faults with hierarchical relationships, the coarse faults are eliminated.

[0043] The specific process of step (3) is as follows:

[0044] (3-1) According to the symptom nodes and fault nodes of the Bayesian network, a BDeu (Bayesian-Dirichlet equivalent uniform prior) function is determined as a scoring function;

[0045] (3-2) Based on the global optimization of the hill climbing search algorithm, the search starts from an initial model. At each step of the search, the search operator is used to locally modify the current model to obtain a series of candidate models;

[0046] (3-3) Use the BDeu function to calculate the score of each candidate model, and compare the best candidate model with the current model. If the score of the best candidate model is higher than the score of the current model, the best candidate model is used as the next model and the search continues. Otherwise, the search is stopped and the model is returned.

[0047] (3-4) Finally, the structure of the Bayesian network with different fault granularities is obtained.

[0048] In step (3-1), the scoring function is as follows:

[0049]

[0050] Where n and m are the number of fault nodes and symptom nodes in the Bayesian network, respectively, and r i Represents any node X i The number of states, q i Represents node X i The number of parent nodes; N ijk Indicates the training set X i The number of samples whose parent node set is in the jth state combination when in the kth state; N ij It is obtained by calculation, which represents the prior distribution of the Bayesian network structure G, and α is a hyperparameter.

[0051] The specific process of step (4) is as follows:

[0052] (4-1) For the faulty node, define and determine the prior probability of each fault state occurring;

[0053] (4-2) For the symptom node, the conditional probability of the symptom node occurring when the fault node occurs is defined by the MLE algorithm (Maximum Likelihood Estimation): for the preprocessed data set D obtained in step (2-1) = {D1, D2, …, D m}, the maximum likelihood estimate l(θ|D) of the conditional probability distribution θ of the symptom node is calculated as:

[0054] l(θ|D)=logP(θ|D)=logΠ l P(θ l |D)=∑ l logP(θ l |D)=∑ ijk N ijk logθ ijk

[0055] Among them, θ ijk =P(X i = k|Pa(X i )=j) represents the set of faulty parent nodes Pa(X i ) is in the jth state combination when the symptom node X i The conditional probability of being in the kth state, N ijk Represents X i The number of samples whose parent node set is in the jth state combination when it is in the kth state; by maximizing l(θ|D), the estimated value of each conditional probability parameter is obtained

[0056]

[0057] Among them, N ij =∑ k N ijk .

[0058] The specific process of step (5) is as follows:

[0059] (5-1) The fault symptoms of the air handling unit are obtained based on sensor measurement data or field observation information and input into the Bayesian network. The evidence will propagate in the network and update the posterior probability of each fault occurrence. The size of the posterior probability represents the possibility of the fault occurrence, thereby helping to determine the most likely type of fault.

[0060] If F i is a fault type, E j is an observed sign, then when the sign appears, it is E j = True, a fault occurs, that is, F i =True can be expressed as:

[0061]

[0062] i=1,2,…,n, where P(F i =True) is called the prior probability, P(E j =True|F i =True) is called conditional probability, P(F i =True|E j =True) is called the posterior probability.

[0063] (5-2) If the posterior probability of a faulty node exceeds the first threshold, an alarm is issued for the fault; if the posterior probability of a faulty node does not exceed the first threshold, but the difference between the posterior probability of a faulty node and the second highest posterior probability exceeds the second threshold, an alarm is issued for the fault. The first threshold is 0.7 and the second threshold is 0.3.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] Compared with traditional fault detection and diagnosis methods, this method has higher versatility and more accurate diagnostic results. When a sensor arranged in an air handling unit fails or is migrated to a building with the same equipment configuration but fewer sensors, this method can automatically adjust the number and level of the Bayesian network fault nodes, and ensure the accuracy and precision of the diagnostic results on the basis of making full use of sensor information; here, accuracy means that the result of the Bayesian network diagnosis is correct, and precision means that on the basis of ensuring accuracy, the fault level diagnosed by the Bayesian network is as high as possible and the granularity is as fine as possible. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1(a) shows the traditional Bayesian network constructed under the condition of sufficient omen information.

[0067] Figure 1(b) shows the traditional Bayesian network constructed under the condition of missing omen information.

[0068] Figure 1(c) shows a Bayesian network with node granularity constructed when symptom information is missing. Figure 1(d) shows a Bayesian network with overlapping faults constructed when symptom information is missing.

[0069] Figure 2 The present invention is a flow chart of automatically constructing a Bayesian network with different node granularities for air handling unit fault detection and diagnosis.

[0070] FIG3(a) is a Bayesian network with node granularity constructed in an embodiment of the present invention.

[0071] FIG. 3( b ) is a traditional Bayesian network constructed in an embodiment of the present invention DETAILED DESCRIPTION

[0072] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0073] The importance of building a Bayesian network with fault classification was mentioned in the background technology section. Here, the necessity of fault classification is introduced in detail. A traditional Bayesian network is constructed. The network consists of two detailed faults: cooling coil valve leakage and cooling circulation pump failure, and five symptoms: supply air temperature, cooling coil power consumption, heating coil power consumption, cooling water circulation pressure difference, and cooling water circulation pump control signal. As shown in Figure 1(a), when the information of the five symptoms is fully determined, each fault can be accurately diagnosed. However, when the sensor arranged in the air handling unit fails or is migrated to a building with the same equipment configuration but a smaller number of sensors, the sensor arrangement changes and the collected symptom information also changes, resulting in the loss of symptom information determined in the Bayesian network. At this time, the diagnosis results of cooling coil valve leakage and chilled water circulation pump status failure become vague, as shown in Figure 1(b); a reasonable way to deal with it is to find Find out the coarse faults corresponding to the two fine faults. After obtaining the approximate location of the air handling unit fault, further targeted detection can be carried out to finally determine the detailed location of the fault. As shown in Figure 1(c), after determining that the coarse fault is insufficient power of the cold water circulation, the cold water circulation power system in the air handling unit can be further detected to finally determine a more accurate fault location. However, it should be noted that in order to avoid the reduction of the accuracy of the Bayesian network diagnosis results, the existence of overlapping faults should be avoided in a Bayesian network. As shown in Figure 1(d), due to the overlap of the coarse fault of insufficient power of the cold water circulation and the two fine faults of cooling coil valve leakage and cold water circulation pump failure, the diagnosis result is still poor, so it is necessary to avoid coarse faults and fine faults appearing in the Bayesian network at the same time. Therefore, it is necessary to select appropriate fault nodes based on the determined symptom information and construct a Bayesian network.

[0074] like Figure 2 As shown, the automatic construction method of Bayesian networks with different node granularities for air handling unit fault detection and diagnosis mainly includes five steps: fault level calculation for air handling unit fault set based on knowledge granularity, determination of nodes of Bayesian networks with different fault granularities, data-driven structure learning of Bayesian networks with different fault granularities, data-driven parameter learning of Bayesian networks with different fault granularities, and air handling unit fault detection and diagnosis using the Bayesian network structure after parameter learning;

[0075] In the first step, the main purpose of calculating the fault level of the air handling unit fault set based on knowledge granularity is to solve the problem of reduced fault levels due to reduced symptom information. Based on fault diagnosis knowledge, it can be deduced that when faced with more symptom information, experts can often interpret more information, thereby obtaining more accurate fault information in diagnosis, that is, high-level faults; when the number of sensors decreases, that is, when the effective symptom information decreases, the information interpreted by experts is also reduced accordingly, so that the fault information obtained in diagnosis is more vague, that is, low-level faults. Through this step, the fault level is divided based on historical data, laying the foundation for the subsequent construction of Bayesian networks with different fault levels.

[0076] In the second step, the nodes of the Bayesian network with different fault granularities are determined. The main purpose is to realize the automatic selection of fault nodes of different granularities based on the richness of fault level information. In actual diagnosis, after obtaining sensor data, the richness of the symptom information corresponding to the system can be obtained (called known symptom information), but which faults can be diagnosed by the known symptom information and the corresponding fault level, and the relationship between the fault and the symptom needs to be further determined. Based on attribute reduction, the minimum symptom information required to diagnose a certain fault (called minimum symptom information) is determined through historical data; when the minimum symptom information is included in the known symptom information, the fault can be diagnosed by the current known symptom information, otherwise it cannot be diagnosed; at the same time, considering the fault level relationship determined in step one, the fault with a higher fault level and can be diagnosed is selected as the fault node of the Bayesian network.

[0077] In the third step, the purpose of data-driven Bayesian network structure learning with different fault granularities is to obtain the corresponding Bayesian network structure with different fault levels by using the scoring search algorithm based on the determination of fault nodes and symptom nodes.

[0078] In the fourth step, the purpose of data-driven Bayesian network parameter learning with different fault granularities is to obtain the node parameters of the Bayesian network using the maximum likelihood estimation algorithm based on the determined Bayesian network structure.

[0079] In the fifth step, the purpose of fault detection and diagnosis of air handling units with different node granularity is to use a fully trained Bayesian network for fault detection and diagnosis, input known symptom information, calculate the posterior probability of each fault to determine the root cause of the abnormal characterization, and take into account the diagnostic accuracy while retaining the highest precision.

[0080] S1: First, based on the knowledge granularity, the fault level and minimum attribute reduction symptom set of the air handling unit fault set are calculated. The specific steps are as follows:

[0081] Step 1-1, based on a hierarchical set of historical data, obtain all possible fault information and all possible symptom information of an air handling unit with a certain equipment configuration, wherein the certain equipment configuration refers to a full understanding of information such as the type and location layout of all equipment in the air handling unit.

[0082] Step 1-2: preprocess the obtained hierarchical set of historical data.

[0083] In steps 1-3, a fault is randomly selected as a decision attribute, and other faults are taken as conditional attributes. The attributes of the selected faults are simplified to obtain the minimum attribute simplified fault set, knowledge granularity, and relative knowledge granularity.

[0084] In steps 1-4, when the relative knowledge granularity of the selected fault is 0, it means that the selected fault and the minimum attribute reduction set have a hierarchical relationship. If the knowledge granularity of the selected fault is smaller than the knowledge granularity of the minimum attribute reduction fault set, then the fault is a coarse fault of the minimum attribute reduction set, otherwise, it is a fine fault.

[0085] In steps 1-5, for all faults, steps 1-3 and 1-4 are repeated to obtain the hierarchical relationship of all faults based on the calculation results.

[0086] In steps 1-6, a certain fault is randomly selected as the decision attribute, all the symptom data in the historical data stratification set are used as conditional attributes, and the attributes of the selected fault are simplified to obtain the minimum attribute simplified symptom set.

[0087] Steps 1-7: For all faults, loop through steps 1-6 to obtain the minimum attribute reduction symptom set for each fault.

[0088] S2: Determine the nodes of the Bayesian network with different fault granularities. The specific steps are as follows:

[0089] According to the fault level sorting and the calculation results of the minimum attribute reduction symptom set, the fault nodes in the Bayesian network are determined. The specific steps are as follows:

[0090] Step 2-1, based on the minimum attribute reduction symptom set and the known situation of the current symptom information, determine the set of faulty nodes that can be diagnosed.

[0091] In step 2-1-1, according to the arrangement of sensors in the air handling unit to be diagnosed, historical data are collected to build a set, the data set is preprocessed, and the known signs and conscription nodes of the Bayesian network are determined.

[0092] In step 2-1-2, a fault is randomly selected, and based on the relationship between the minimum attribute reduction symptom set of the fault and the known symptom set, whether to select the node in the Bayesian network is determined: if the minimum attribute reduction symptom set of the fault is a subset of the known symptom set, then the fault is introduced into the Bayesian network fault node set, otherwise it is not introduced.

[0093] In step 2-1-3, loop step 2-1-2, traverse all faults, and obtain the set of fault nodes constructed by the Bayesian network.

[0094] Step 2-2, based on the fault hierarchy sorting, filter out overlapping faults: if the introduced faults contain both coarse faults and all fine faults with hierarchical relationships, the coarse faults will be eliminated.

[0095] S3: Data-driven structure learning of Bayesian networks with different fault granularities. The specific steps are:

[0096] Step 3-1, determining the symptom nodes and fault nodes of the Bayesian network according to step S2.

[0097] Step 3-2, determine the scoring function. In this strategy, the Bayesian network structure is constructed through the scoring search algorithm. The BDeu function is used as the scoring function in this model, so the scoring function is determined as:

[0098]

[0099] Where n and m are the number of fault nodes and symptom nodes in the Bayesian network, respectively, and r i Represents any node X i The number of states, q i Represents node X i The number of parent nodes. N ijk Indicates the training set X i The number of samples whose parent node set is in the jth state combination when in the kth state. N ij It can be obtained by calculation. It represents the prior distribution of the Bayesian network structure G, and α is a hyperparameter.

[0100] Step 3-3, global optimization based on search algorithm:

[0101] In step 3-3-1, the search algorithm used in this method is the hill climbing search algorithm, which starts the search from an initial model (the initial model is generally defined as an unbounded model).

[0102] Step 3-3-2: At each step of the search, the search operator is used to locally modify the current model to obtain a series of candidate models.

[0103] In step 3-3-3, based on the preprocessed data set obtained in step 2-1-1, the scoring function determined in step 3-2 is used to calculate the score for each candidate model, and the optimal candidate model is compared with the current model. If the score of the optimal candidate model is higher than the score of the current model, the optimal candidate model is used as the next model and the search continues. Otherwise, the search is stopped and the model is returned.

[0104] Finally, the structure of the Bayesian network with different fault granularities is obtained.

[0105] S4: Data-driven parameter learning of Bayesian networks with different fault granularities. The specific steps are:

[0106] Step 4-1: In this method, for the faulty node, the prior probability of each fault state is defined. This value can be estimated through literature research, field research, and historical system operation and maintenance records.

[0107] Step 4-2: For the symptom node, the conditional probability of the symptom node occurring when the fault node occurs is defined by the MLE algorithm: for the preprocessed data set D obtained in step 2-1-1, D = {D1, D2, ..., D m}, the maximum likelihood estimate l(θ|D) of the conditional probability distribution θ of the symptom node can be calculated as:

[0108] l(θ|D)=logP(θ|D)=logΠ l P(θ l |D)=∑ l logP(θ l |D)=∑ ijk N ijk logθ ijk

[0109] Among them, θ ijk =P(X i = k|Pa(X i )=j) represents the set of faulty parent nodes Pa(X i ) is in the jth state combination when the symptom node X i The conditional probability of being in the kth state, N ijk Represents X i The number of samples whose parent node set is in the jth state combination when in the kth state. By maximizing l(θ|D), the estimated value of each conditional probability parameter can be obtained.

[0110]

[0111] Among them, N ij =Σ k N ijk .

[0112] S5: Use the Bayesian network structure after parameter learning to perform fault detection and diagnosis of air handling units. The specific steps are as follows:

[0113] In step 5-1, the fault symptoms of the air handling unit are obtained based on sensor measurement data or field observation information and input into the Bayesian diagnostic network. The evidence will propagate in the network and update the posterior probability of each fault occurrence. The size of the evidence represents the possibility of the fault occurrence, thereby helping to determine the most likely type of fault.

[0114] If F i is a fault type, E j is an observed sign, then when the sign appears, it is E j = True, a fault occurs, that is, F i =True can be expressed as:

[0115]

[0116] i=1,2,…,n, where P(F i =True) is called the prior probability, P(E j =True|F i =True) is called conditional probability, P(F i =True|E j =True) is called the posterior probability.

[0117] In step 5-2, if the posterior probability of a faulty node exceeds the first threshold, an alarm is issued for the fault; if the posterior probability of a faulty node does not exceed the first threshold, but the difference with the second highest posterior probability exceeds the second threshold, an alarm is issued for the fault. The first threshold is 0.7 and the second threshold is 0.3.

[0118] Based on the above technical solution, the following preferred embodiments can be further provided, wherein:

[0119] The historical data hierarchical set is a more comprehensive data set, which should contain as many fault and symptom features as possible; the symptom features of the historical data construction set are part of the historical data hierarchical set, mainly due to sensor failure or insufficient number of sensors installed in new buildings, resulting in a decrease in the number of symptoms.

[0120] In step S1, the minimum attribute reduction aims to obtain the decision or classification rules of the problem by removing redundant attributes while keeping the classification ability of the data set unchanged. The specific steps are as follows:

[0121] Step 1: Determine the decision table S =<U,C∪D,V,f> , where U is the domain, C and D are conditional attributes and decision attribute sets respectively, V is the range, f is an information function, and an empty reduction set RED is defined at the same time U ;

[0122] Step 2: Select any attribute a from the conditional attribute set C j , calculate the property a j On the importance of internal attributes of condition attribute set C relative to decision attribute set D

[0123]

[0124] if Then a j Add to RED U In the above example, we traverse all the attributes in C and obtain the reduced set RED that is initially screened. U_B ;

[0125] Step 3, if the reduction set RED U_B It has different partitioning capabilities from the total condition attribute set C, that is, GP U (D|RED U_B )≠GP U (D|C); Subtract the reduction set RED from the condition attribute set C U_B Get the attribute set C-RED U_B , from the attribute set C-RED U_B Pick any attribute a j , calculate the property a j About the conditional attribute set RED U_B The importance of external attributes relative to the decision attribute set D

[0126]

[0127] Traversing the attribute set C-RED U_B Among all the attributes in , select the attribute a0 with the greatest external attribute importance and add it to RED U_B middle;

[0128] Step 4: Repeat step 3 until the reduced set RED is obtained. U_B It has the same partitioning capability as the set of all conditional attributes C;

[0129] Step 5. From the new attribute set RED U_B Any attribute a j Get the attribute set RED U_B -{a j}, if the attribute set RED U_B-{a j} has the same partitioning ability as the entire condition attribute set C, then a j From the attribute set RED U_B Delete from a and loop through a j , and obtain the final minimum attribute reduction set.

[0130] In steps 1-3, the calculation formulas for knowledge granularity and relative knowledge granularity are:

[0131]

[0132] GP U (D|B)=GP U (B)-GP U (B∪D)

[0133] In the formula, GP U (B) is the knowledge granularity GP of set B U (B), GP U (D|B) is the relative knowledge granularity of set D relative to set B. Given a decision table S = (U, A = C∪D, V, f), U is the domain, if U / B={X1,X2,...,X m}, there are m equivalence classes, |X i | represents the number of elements in the i-th equivalence class, where i = 1, 2…m.

[0134] In steps 1-2, data preprocessing includes data cleaning and transformation to eliminate abnormal data.

[0135] Example

[0136] In this embodiment, the ASHRAE RP1312 summer data set is selected for demonstration. The fault node is represented by the letter F, and the symptom node is represented by the letter E, as shown in Table 1 and Table 2 below. A total of 14 symptoms and 13 faults are selected.

[0137] Table 1 Faulty nodes

[0138]

[0139] According to the symptom data, there is a fault hierarchy relationship between the 13 faults:

[0140] F10 External air circulation abnormality is a coarse-grained fault of F1 exhaust valve stuck (fully open) and F4 fresh air valve stuck (fully closed);

[0141] The abnormal control of the airflow organization of F11 AHU is a coarse-grained fault of the F2 exhaust valve stuck (fully closed) and the F8 AHU air duct leakage (after the blower);

[0142] The F12 coil valve fault is a coarse-grained fault of the F3 cooling coil valve reversing and the F6 heating coil valve leaking (stage 1-0.4GPM);

[0143] The leakage at the front end of F13AHU is a coarse-grained fault of leakage of F7 fresh air valve (opening degree 45%) and leakage of F9 AHU air duct (before the blower);

[0144] That is, there are 9 fine-grained faults F1-F9 in total, and 4 coarse-grained faults F10-F13 in total.

[0145] Table 2 Symptom nodes

[0146]

[0147] Using Python, the RP1312-summer data is randomly divided into a historical data stratification set and a historical data construction set in a ratio of 7:3. The historical data stratification set is used to obtain the fault hierarchy relationship and the minimum attribute reduction symptom set, and the historical data construction set is used to build a Bayesian network.

[0148] In this embodiment, based on the hierarchical set of historical data, the fault level sorting strategy is implemented according to the aforementioned S1 to obtain the minimum attribute simplified fault set, knowledge granularity and relative knowledge granularity of all faults, as shown in Table 3.

[0149] Table 3 Fault-fault attribute reduction classification results

[0150]

[0151] According to steps 1-4 of S1, the fault level is determined. Taking F1 as an example, when the relative knowledge granularity of F1 is 0, it means that F1 and the minimum attribute simplification fault set {F4, F10} have a hierarchical relationship; and GP(F1)=0.954>max{GP(F4),GP(F10)}=max{0.760,0.727}=0.760, then F1 is a fine fault of the minimum attribute simplification fault set {F4, F10}; taking F13 as an example, when the relative knowledge granularity of F13 is 0, it means that F13 and the minimum attribute simplification fault set {F7, F9} have a hierarchical relationship; and GP(F13)=0.688>max{GP(F7),GP(F9)}=min{0.903,0.756}=0.756, then F13 is a coarse fault of the minimum attribute simplification fault set {F7, F9}.

[0152] In this embodiment, based on the hierarchical set of historical data, the minimum attribute reduction symptom set calculation strategy is implemented according to the aforementioned S1 to obtain the minimum attribute reduction symptom set of all faults, as shown in Table 4.

[0153] Table 4 Fault-symptom attribute reduction results

[0154]

[0155] Due to the lack of sensor data, only part of the symptom data is collected under the same AHU configuration in different buildings. According to S2, according to the arrangement of sensors in the air handling unit to be diagnosed, historical data is collected to determine the known symptom set of the Bayesian network as {E1, E2, E3, E4, E5, E6, E10}. According to the minimum attribute reduction symptom set obtained in Table 4, compared with the known symptom set, it can be determined that the candidate fault nodes are {F2, F3, F8, F9, F10, F11}. According to step 2-2 of S2, since F11 and {F2, F8} exist at the same time, and F11 is a coarse fault node with a hierarchical relationship of {F2, F8}, F11 is eliminated, and the fault nodes finally introduced into the network are {F2, F3, F8, F9, F10}. It should be noted that F10 is a coarse fault node of {F1, F4}, indicating that the current symptom information is not sufficient to support the detection and diagnosis of the fine fault node {F1, F4}.

[0156] According to S3, S4, and S5, a Bayesian network is constructed for fault detection and diagnosis. The network structure is shown in Figure 3(a). The diagnostic accuracy is compared with that of the traditional Bayesian network fault diagnosis method without fault level sorting and screening (the network structure is shown in Figure 3(b)). The results are shown in Table 5.

[0157] Table 5 Diagnostic accuracy of the fault diagnosis method proposed by the present invention and the traditional method

[0158]

[0159] The present invention can perform fault detection and diagnosis on air handling units with different sensor configurations, and does not limit the types and numbers of faults and symptoms. For example, in this embodiment, seven symptoms such as abnormal opening of the E1 heating coil valve and abnormal opening of the E2 cooling coil valve are used for verification, and other symptom nodes are selected to form a known symptom set. The level and diagnostic accuracy of the fault node selection are similar to those of this embodiment, so the present invention is still applicable.

[0160] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for fault detection and diagnosis of air handling units with different node granularities, characterized in that: The following steps are involved: (1) Based on knowledge granularity, the fault level and minimum attribute reduction symptom set of the fault set of the air handling unit are calculated; The specific process is: (1-1) Obtain a hierarchical set of historical data, including all fault information and symptom information of the air handling unit; (1-2) Preprocessing the obtained hierarchical set of historical data; (1-3) Randomly select a fault as the decision attribute, take other faults as conditional attributes, perform minimum attribute simplification on the selected fault, and obtain the minimum attribute simplification fault set, knowledge granularity, and relative knowledge granularity; The calculation formulas for knowledge granularity and relative knowledge granularity are: GP U (D|B)=GP U (B)-GP U (B∪D) In the formula, GP U (B) is the knowledge granularity of set B, GP U (D|B) is the relative knowledge granularity of set D relative to set B, GP U (B∪D) is the knowledge granularity of the union of sets B and D; given a decision table S = (U, A = C∪D, V, f), U is the domain, if U / B={X1,X2,...,X m }, there are m equivalence classes, |X i | represents the number of elements in the i-th equivalence class, where i = 1, 2…m; (1-4) When the relative knowledge granularity of the selected fault is 0, it means that the selected fault and the minimum attribute reduction fault set have a hierarchical relationship. If the knowledge granularity of the selected fault is smaller than the knowledge granularity of the minimum attribute reduction fault set, then the fault is a coarse fault of the minimum attribute reduction fault set, otherwise, it is a fine fault. (1-5) For all faults, loop through steps (1-3) and (1-4) to obtain the hierarchical relationship of all faults based on the calculation results; (1-6) Randomly select a fault as the decision attribute, use all the symptom data in the historical data layer set as the conditional attributes, perform minimum attribute simplification on the selected fault, and obtain the minimum attribute simplified symptom set; (1-7) For all faults, loop through steps (1-6) to obtain the minimum attribute reduction symptom set for each fault; (2) Determine the symptom nodes and fault nodes in the Bayesian network based on the calculation results of the fault level sorting and the minimum attribute reduction symptom set; (3) Based on the symptom nodes and fault nodes of the Bayesian network, a Bayesian network structure is constructed through a scoring search algorithm to obtain a Bayesian network structure with different fault granularities; (4) Perform data-driven parameter learning on the constructed Bayesian network structure; (5) The Bayesian network structure after parameter learning is used to detect and diagnose air handling unit faults. The abnormal symptoms of the air handling unit are input and the posterior probability of each fault occurring is output.

2. The method for detecting and diagnosing air handling unit faults with different node granularities according to claim 1, characterized in that: The specific process of minimum attribute reduction is: Step 1: Determine the decision table S =<U,C∪D,V,f> , where U is the domain, C and D are the condition attribute set and decision attribute set respectively, V is the range, f is an information function, and an empty reduction set RED is defined at the same time U ; Step 2: Select any attribute a from the conditional attribute set C j , calculate the property a j On the importance of internal attributes of condition attribute set C relative to decision attribute set D if Then a j Add to RED U In the above example, we traverse all the attributes in C and obtain the reduced set RED that is initially screened. U_B ; Step 3, if the reduction set RED U_B It has different partitioning capabilities from the total condition attribute set C, that is, GP U (D|RED U_B )≠GP U (D|C); Subtract the reduction set RED from the condition attribute set C U_B Get the attribute set C-RED U_B , from the attribute set C-RED U_B Pick any attribute a j , calculate the property a j About the conditional attribute set RED U_B The importance of external attributes relative to the decision attribute set D Traversing the attribute set C-RED U_B Among all the attributes in , select the attribute a0 with the greatest external attribute importance and add it to RED U_B middle; Step 4: Repeat step 3 until the reduced set RED is obtained. U_B It has the same partitioning capability as the set of all conditional attributes C; Step 5. From the new attribute set RED U_B Select any attribute a j Get the attribute set RED U_B -{a j }, if the attribute set RED U_B -{a j } has the same partitioning ability as the entire condition attribute set C, then a j From the attribute set RED U_B Delete from a and loop through a j , and obtain the final minimum attribute reduction set.

3. The method for detecting and diagnosing air handling unit faults with different node granularities according to claim 1, characterized in that: The specific process of step (2) is as follows: (2-1) According to the arrangement of sensors in the air handling unit to be diagnosed, historical data are collected to build a set, the data set is preprocessed, and the known signs and sign nodes of the Bayesian network are determined; (2-2) Randomly select a fault and determine whether to select the node in the Bayesian network based on the relationship between the minimum attribute reduction symptom set of the fault and the known symptom set: if the minimum attribute reduction symptom set of the fault is a subset of the known symptom set, then introduce the fault into the Bayesian network fault node set, otherwise do not introduce it; (2-3) Loop through step (2-2), traverse all faults, and obtain the set of fault nodes constructed by the Bayesian network; (2-4) Based on the fault hierarchy sorting, overlapping faults are screened out: if the introduced faults contain both coarse faults and all fine faults with hierarchical relationships, the coarse faults are eliminated.

4. The method for detecting and diagnosing air handling unit faults with different node granularities according to claim 1, characterized in that: The specific process of step (3) is as follows: (3-1) According to the symptom nodes and fault nodes of the Bayesian network, the BDeu function is determined as the scoring function; (3-2) Based on the global optimization of the hill climbing search algorithm, the search starts from an initial model. At each step of the search, the search operator is used to locally modify the current model to obtain a series of candidate models; (3-3) Use the BDeu function to calculate the score of each candidate model, and compare the best candidate model with the current model. If the score of the best candidate model is higher than the score of the current model, the best candidate model is used as the next model and the search continues. Otherwise, the search is stopped and the model is returned. (3-4) Finally, the structure of the Bayesian network with different fault granularities is obtained.

5. The method for detecting and diagnosing air handling unit faults with different node granularities according to claim 4, characterized in that: In step (3-1), the scoring function is as follows: Where n and m are the number of fault nodes and symptom nodes in the Bayesian network, respectively, and r i Represents any node X i The number of states, q i Represents node X i The number of parent nodes; N ijk Indicates the training set X i The number of samples whose parent node set is in the jth state combination when in the kth state; N ij By N ijk It is calculated that P(G) represents the prior distribution of the Bayesian network structure G, and α is a hyperparameter.

6. The method for detecting and diagnosing air handling unit faults with different node granularities according to claim 1, characterized in that: The specific process of step (4) is as follows: (4-1) For the faulty node, define and determine the prior probability of each fault state occurring; (4-2) For the symptom node, the conditional probability of the symptom node occurring when the fault node occurs is defined by the MLE algorithm: for the preprocessed data set D obtained in step (2-1) = {D1, D2, …, D m }, the maximum likelihood estimate l(θ|D) of the conditional probability distribution θ of the symptom node is calculated as: Among them, θ ijk =P(X i = k|Pa(X i )=j) represents the set of faulty parent nodes Pa(X i ) is in the jth state combination when the symptom node X i The conditional probability of being in the kth state, N ijk Represents X i The number of samples whose parent node set is in the jth state combination when in the kth state; by maximizing l(θ|D), the estimated value of each conditional probability parameter is obtained Among them, N ij =∑ k N ijk .

7. The method for detecting and diagnosing air handling unit faults with different node granularities according to claim 1, characterized in that: The specific process of step (5) is as follows: (5-1) The fault symptoms of the air handling unit are obtained based on sensor measurement data or field observation information and input into the Bayesian network. The fault symptoms will propagate in the network and update the posterior probability of each fault occurrence. The size of the posterior probability represents the possibility of the fault occurrence, thereby helping to determine the most likely type of fault; If F i is a fault type, E j is an observed sign, then when the sign appears, it is E j = True, a fault occurs, that is, F i =True can be expressed as: i=1,2,…,n, where P(F i =True) is called the prior probability, P(E j =True|F i =True) is called conditional probability, P(F i =True|E j =True) is called the posterior probability; (5-2) If the posterior probability of a faulty node exceeds the first threshold, it is diagnosed as the fault; if the posterior probability of a faulty node does not exceed the first threshold, but the difference with the second highest posterior probability exceeds the second threshold, it is diagnosed as the fault; the first threshold is 0.7 and the second threshold is 0.3.

Images