Residual-knowledge-data joint driven fault composite diagnosis method and system

Abstract

A residual-knowledge-data jointly driven fault composite diagnosis method and system, in which the first step is to construct a probability graph model, including determining the structure and parameters of the probability graph model, and the second step is to apply the probability graph model online, calculate the online real-time monitoring data, and input the residual data of the features in the residual nodes, the direct monitoring data of the features in the data nodes, and the state of the features in the knowledge nodes into the probability graph model to obtain the posterior probability of each state of the diagnosis layer node, and output the state with the maximum posterior probability in the diagnosis layer as the fault diagnosis result; the system includes a memory and a processor, and the processor executes a computer program to implement the above method. The present invention integrates the advantages of the three types of methods, namely, residual-driven, knowledge-driven, and data-driven, into the same diagnostic system, realizes the complementary advantages of each other, and obtains better diagnostic performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular relates to a residual-knowledge-data jointly driven fault composite diagnosis method and system. Background Art

[0002] In recent years, building energy consumption has increased year by year, and currently accounts for 40% of the total energy consumption of social terminals. Air conditioning and refrigeration systems are the main energy consumption sources in buildings. And chillers are the main energy-consuming equipment in air conditioning and refrigeration systems. When a chiller fails, if it is not eliminated in time, it will lead to a decrease in the energy efficiency of the unit, resulting in a large amount of energy waste, reducing the service life of the equipment, and reducing the comfort of the indoor environment. The operation of the chiller with a fault will increase energy consumption by up to 30%. Therefore, the application of fault diagnosis technology is of great significance to eliminate chiller failures and achieve energy saving.

[0003] Nowadays, the fault diagnosis of chillers has been widely studied, and a large number of fault diagnosis methods for chillers have been proposed. Classically, these methods are divided into three categories: model-driven methods, knowledge-driven methods, and data-driven methods. In fact, each single-category method has its own unique advantages. The main advantage of the model-driven method is that the physical model is often based on a deep understanding of the system and the fault mechanism, which can well reflect the system characteristics, and the model construction does not require a large amount of data. The most typical application of model-driven is to use the model to determine the benchmark value of the characteristic parameter operation, and then obtain the characteristic residual by comparing the benchmark value with the measured value. At this time, the model-driven can also be called residual-driven. The main advantage of the knowledge-driven method is that it does not require the construction of an accurate mechanism model, the reasoning mechanism is clear and transparent, and a large number of accumulated expert rules and prior knowledge are fully utilized. The main advantage of the data-driven method is that it does not require the construction of an accurate mechanism model, nor does it require a large amount of expert knowledge. It is very suitable for occasions where it is impossible to build an accurate physical model and lacks detailed prior knowledge. These advantages are unique to the method of the category to which they belong, resulting in different advantages for different methods. Moreover, these methods also have obvious disadvantages. For example, when it is difficult to build an effective benchmark value model, the residual-driven and knowledge-driven methods will not be able to complete the diagnosis work; when effective diagnostic rules cannot be established, the knowledge-driven method will not be able to work independently; when it is difficult to obtain enough fault data, the residual-driven and data-driven methods will not be able to work independently. Obviously, it is necessary to build a composite diagnostic method that integrates the advantages of different types of methods, but there are also certain difficulties. Summary of the invention

[0004] The purpose of the present invention is to provide a residual-knowledge-data jointly driven composite fault diagnosis method and system, which integrates the advantages of residual-driven, knowledge-driven and data-driven methods into the same diagnostic system, and allows them to work in parallel, to construct a composite diagnosis method jointly driven by these three types of models, and to achieve complementary advantages.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is: a residual-knowledge-data jointly driven fault composite diagnosis method, comprising the following steps:

[0006] (1) obtaining data of the chiller under normal and faulty operating conditions, filtering abnormal and non-steady-state data in the normal and faulty data in a steady state, obtaining steady-state data of the chiller under normal and faulty operating conditions, and selecting an indicator feature for determining the operating condition of the chiller;

[0007] (2) Determine the structure of the probabilistic graphical model; the structure of the probabilistic graphical model includes three layers, from top to bottom, namely, the diagnosis layer, the fault layer and the symptom layer, wherein the diagnosis layer is a discrete node D with n+1 states, representing n known fault states and 1 new fault state respectively; the fault layer includes n fault nodes, each fault node represents a specific known fault F i , there are two states, "True" and "False", which represent the known fault F i Occurrence and non-occurrence; in the symptom layer, each known fault F i They are connected to three types of nodes, namely, a residual node R, a data node X and several knowledge nodes K. The residual node R and the data node X are m-dimensional continuous nodes composed of m features representing the health status of the chiller, and their data obey the m-dimensional Gaussian distribution. Each knowledge node K represents a feature, and there are three states representing that the measured value of the feature is "significantly increased", "basically unchanged" and "significantly decreased" relative to the baseline value of the feature when the chiller is operating normally; the specific type of the knowledge node K is determined by the association rule between the fault and the feature;

[0008] (3) Determine the parameters of the probability graph model; the parameters of the probability graph model include the prior probabilities of the n+1 states of the diagnosis layer node D, the known faults F in the fault layer, i The conditional probabilities of the two states, the conditional probability distribution of the residual node R and the data node X in the symptom layer, and the conditional probability of the knowledge node K in the symptom layer; the prior probability of the n+1 states of the diagnosis layer node D is obtained by expert experience or statistical unit history failure frequency; the known fault F in the fault layer i The conditional probabilities of the two states are assigned according to the following three rules:

[0009] Rule 1: P(Fi =True|D=F i )=1,P(F i =False|D=F i )=0;

[0010] Rule 2: P(F i =True|D=F j )=0,P(F i =False|D=F j )=1,i≠j;

[0011] Rule 3: P(F i =True|D=NF)=0,P(F i =False|D=NF)=1; where NF indicates a new fault.

[0012] The conditional probability distribution of the residual node R in the symptom layer is given by the known fault F i When the residual data of each feature is in the "True" state, it is obtained by maximum likelihood estimation; the conditional probability distribution of the data node X in the symptom layer is determined by the known fault F i The direct monitoring steady-state data of each feature when it is in the "True" state is obtained by maximum likelihood estimation; the conditional probability of the knowledge node K in the symptom layer under the three states is obtained by expert experience or historical data statistics under fault operating conditions;

[0013] (4) Selecting a reference value model for determining the reference value of the feature in the knowledge node K, taking the features constituting the knowledge node K as the output of the model, and taking the indicative features of the chiller operating conditions as the input of the model, and using the acquired steady-state data under normal operating conditions of the chiller to train the reference value model, thereby obtaining a constructed reference value model;

[0014] (5) Selecting a feature change significance judgment method, judging whether the feature in the knowledge node K has changed significantly relative to the reference value, and determining the judgment threshold of the feature change significance judgment method;

[0015] (6) Real-time data of the chiller during operation is obtained. After steady-state filtering and feature selection, the residual data of the features in the residual node R in the symptom layer, the direct monitoring data of the features in the data node X, and the status of the features in the knowledge node K are obtained. These data are input into the constructed probabilistic graphical model as evidence, and the posterior probability of each state of the diagnosis layer node D is calculated to determine what fault has occurred in the chiller.

[0016] In step (2), for a known fault F iThe specific type of knowledge node K is determined by the association rules between faults and features. The specific method is as follows: first, based on the principle of thermodynamics, the thermodynamic change law between faults and features is analyzed to determine the known fault F i and the association rules between the features; then all the i Features with strong thermodynamic relationships are identified as known faults F i The connected knowledge nodes, the strong thermodynamic relationship is that the magnitude of the change in the fault-induced characteristic exceeds a preset threshold.

[0017] The step (3) is to determine the known fault F i When it is the residual data of the feature in the "True" state, it includes the following steps:

[0018] (a) Selecting a reference value model for determining a characteristic reference value, taking the characteristics constituting the residual node R as the output of the model, and taking the indicative characteristics of the chiller operating condition as the input of the model;

[0019] (b) using the acquired steady-state data of the chiller under normal operating conditions to train the benchmark value model to obtain a trained feature benchmark value model;

[0020] (c) Use the acquired steady-state data of the chiller under faulty operating conditions and input it into the trained feature benchmark value model. For each feature, compare the benchmark value output by the model with the actual measured value to obtain the residual data of the two, which is the residual data of the feature.

[0021] In the step (6), the method for obtaining the residual data of the feature in the residual node R in the symptom layer is: inputting the acquired real-time data of the chiller operation into the trained feature benchmark value model, and for each feature, comparing the benchmark value output by the model with the actual measured value to obtain the residual data of the two; the direct monitoring data of the feature in the data node X is the acquired real-time data of the chiller operation; the method for obtaining the state of the feature in the knowledge node K is: inputting the acquired real-time data of the chiller operation into the trained feature benchmark value model to obtain the benchmark value output by the model, and for each feature, based on the feature change significance judgment method and the determined judgment threshold, according to the benchmark value output by the feature benchmark value model and the actual measured value, obtain the judgment results of "significantly increased", "basically unchanged" and "significantly decreased" of the feature relative to the benchmark value, thereby forming the state of the feature in the knowledge node K.

[0022] The residual-knowledge-data jointly driven fault composite diagnosis system proposed in the present invention comprises a memory and a processor, wherein the memory stores a computer program for execution on the processor; when the processor executes the computer program, the following steps are implemented:

[0023] (1) obtaining data of the chiller under normal and faulty operating conditions, filtering abnormal and non-steady-state data in the normal and faulty data in a steady state, obtaining steady-state data of the chiller under normal and faulty operating conditions, and selecting an indicator feature for determining the operating condition of the chiller;

[0024] (2) Determine the structure of the probabilistic graphical model; the structure of the probabilistic graphical model includes three layers, from top to bottom, namely, the diagnosis layer, the fault layer and the symptom layer, wherein the diagnosis layer is a discrete node D with n+1 states, representing n known fault states and 1 new fault state respectively; the fault layer includes n fault nodes, each fault node represents a specific known fault F i , there are two states, "True" and "False", which represent the known fault F i Occurrence and non-occurrence; in the symptom layer, each known fault F i They are connected to three types of nodes, namely, a residual node R, a data node X and several knowledge nodes K. The residual node R and the data node X are m-dimensional continuous nodes composed of m features representing the health status of the chiller, and their data obey the m-dimensional Gaussian distribution. Each knowledge node K represents a feature, and there are three states representing that the measured value of the feature is "significantly increased", "basically unchanged" and "significantly decreased" relative to the baseline value of the feature when the chiller is operating normally; the specific type of the knowledge node K is determined by the association rule between the fault and the feature;

[0025] (3) Determine the parameters of the probability graph model; the parameters of the probability graph model include the prior probabilities of the n+1 states of the diagnosis layer node D, the known faults F in the fault layer, i The conditional probabilities of the two states, the conditional probability distribution of the residual node R and the data node X in the symptom layer, and the conditional probability of the knowledge node K in the symptom layer; the prior probability of the n+1 states of the diagnosis layer node D is obtained by expert experience or statistical unit history failure frequency; the known fault F in the fault layer i The conditional probabilities of the two states are assigned according to the following three rules:

[0026] Rule 1: P(F i =True|D=F i )=1,P(F i =False|D=F i )=0;

[0027] Rule 2: P(F i =True|D=F j )=0,P(F i =False|D=F j )=1,i≠j;

[0028] Rule 3: P(F i =True|D=NF)=0,P(F i =False|D=NF)=1; where NF indicates a new fault.

[0029] The conditional probability distribution of the residual node R in the symptom layer is given by the known fault F i When the residual data of each feature is in the "True" state, it is obtained by maximum likelihood estimation; the conditional probability distribution of the data node X in the symptom layer is determined by the known fault F i The direct monitoring steady-state data of each feature when it is in the "True" state is obtained by maximum likelihood estimation; the conditional probability of the knowledge node K in the symptom layer under the three states is obtained by expert experience or historical data statistics under fault operating conditions;

[0030] (4) Selecting a reference value model for determining the reference value of the feature in the knowledge node K, taking the features constituting the knowledge node K as the output of the model, and taking the indicative features of the chiller operating conditions as the input of the model, and using the acquired steady-state data under normal operating conditions of the chiller to train the reference value model, thereby obtaining a constructed reference value model;

[0031] (5) Selecting a feature change significance judgment method, judging whether the feature in the knowledge node K has changed significantly relative to the reference value, and determining the judgment threshold of the feature change significance judgment method;

[0032] (6) Real-time data of the chiller during operation is obtained. After steady-state filtering and feature selection, the residual data of the features in the residual node R in the symptom layer, the direct monitoring data of the features in the data node X, and the status of the features in the knowledge node K are obtained. These data are input into the constructed probabilistic graphical model as evidence, and the posterior probability of each state of the diagnosis layer node D is calculated to determine what fault has occurred in the chiller.

[0033] In step (2), for a known fault F i The specific type of knowledge node K is determined by the association rules between faults and features. The specific method is as follows: first, based on the principle of thermodynamics, the thermodynamic change law between faults and features is analyzed to determine the known fault F i and the association rules between the features; then all the i Features with strong thermodynamic relationships are identified as known faults F i The connected knowledge nodes, the strong thermodynamic relationship is that the magnitude of the change in the fault-induced characteristic exceeds a preset threshold.

[0034] The step (3) is to determine the known fault F iWhen it is the residual data of the feature in the "True" state, it includes the following steps:

[0035] (a) Selecting a reference value model for determining a characteristic reference value, taking the characteristics constituting the residual node R as the output of the model, and taking the indicative characteristics of the chiller operating condition as the input of the model;

[0036] (b) using the acquired steady-state data of the chiller under normal operating conditions to train the benchmark value model to obtain a trained feature benchmark value model;

[0037] (c) Use the acquired steady-state data of the chiller under faulty operating conditions and input it into the trained feature benchmark value model. For each feature, compare the benchmark value output by the model with the actual measured value to obtain the residual data of the two, which is the residual data of the feature.

[0038] In the step (6), the method for obtaining the residual data of the feature in the residual node R in the symptom layer is: inputting the acquired real-time data of the chiller operation into the trained feature benchmark value model, and for each feature, comparing the benchmark value output by the model with the actual measured value to obtain the residual data of the two; the direct monitoring data of the feature in the data node X is the acquired real-time data of the chiller operation; the method for obtaining the state of the feature in the knowledge node K is: inputting the acquired real-time data of the chiller operation into the trained feature benchmark value model to obtain the benchmark value output by the model, and for each feature, based on the feature change significance judgment method and the determined judgment threshold, according to the benchmark value output by the feature benchmark value model and the actual measured value, obtain the judgment results of "significantly increased", "basically unchanged" and "significantly decreased" of the feature relative to the benchmark value, thereby forming the state of the feature in the knowledge node K.

[0039] The beneficial effects of the present invention are as follows: based on the probabilistic graphical model, the present invention integrates the advantages of the three types of methods, namely, residual-driven, knowledge-driven and data-driven, into the same diagnostic system, and constructs a residual-knowledge-data jointly driven composite diagnostic method and system, thereby realizing the complementary advantages of the three types of methods and obtaining better diagnostic performance, which specifically includes the following points:

[0040] 1) When the composite diagnostic method provided diagnoses faults, the residual, knowledge and data nodes involved in the integration work in parallel, which enables the composite diagnostic method to not only obtain better diagnostic performance than any type of node when the residual-knowledge-data nodes work together, but also tolerate the absence of any type of node in the diagnostic process. Compared with the mechanism of integration in series, the above advantages are more significant. For example, if the residual, knowledge and data nodes are integrated in series, the output of the previous node is the input of the next node, and the rings are connected, the diagnostic performance may be improved, but the composite model in series cannot tolerate the absence of any type of node, otherwise the model will not be able to complete the diagnostic task, and it increases the difficulty of model construction.

[0041] 2) The composite diagnostic method provided fully inherits the advantages of the three types of methods, namely residual, knowledge and data, and makes up for their disadvantages to a certain extent. First, when it is difficult to build an effective benchmark value model, the residual and knowledge-driven methods will not be able to work independently, while the composite diagnostic method provided can only use data nodes for fault diagnosis; second, when effective diagnostic rules cannot be established, the knowledge-driven method will not be able to work independently, while the composite diagnostic method provided can jointly use residual nodes and data nodes for fault diagnosis; third, when it is difficult to obtain enough fault data, the residual nodes and data nodes will not be able to work independently, while the composite diagnostic method provided can only use knowledge nodes for fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flow chart is constructed for the probability graph model of the residual-knowledge-data jointly driven fault composite diagnosis method of the present invention.

[0043] Figure 2 The present invention provides a flowchart for online application of the probability graph model of the residual-knowledge-data jointly driven fault composite diagnosis method.

[0044] Figure 3 It is the general structure of the probability graph model in the residual-knowledge-data jointly driven fault composite diagnosis method described in the present invention.

[0045] Figure 4 , which is the structure of the probability graph model in the embodiment.

[0046] Figure 4 Chinese: F-1 indicates a cooling water volume reduction fault, F-2 indicates a chilled water volume reduction fault, F-3 indicates a refrigerant undercharge fault, F-4 indicates a refrigerant overcharge fault, F-5 indicates a condenser scaling fault, F-6 indicates an excess oil fault, and F-7 indicates a non-condensable gas fault. DETAILED DESCRIPTION

[0047] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the invention in any way.

[0048] Example 1: Refer to the attached Figure 1 and attached Figure 2 As shown, taking the fault diagnosis of a chiller as an example, the specific implementation steps of the fault composite diagnosis method based on residual-knowledge-data joint drive provided by the present invention are as follows:

[0049] Step 1: Probabilistic graphical model construction

[0050] 1a) Data preprocessing, including the following steps:

[0051] 1a-1) Obtain the data of the chiller under normal and fault operating conditions through experiments or historical database stored on site;

[0052] 1a-2) Steady-state filtering: using an existing steady-state filtering method to filter the abnormal and non-steady-state data in the normal and fault data obtained in step 1a-1), to obtain steady-state data of the chiller under normal and fault operating conditions;

[0053] 1a-3) For normal and each fault condition, randomly select no less than 100 groups of steady-state data obtained in step 1a-2) to form a training set;

[0054] 1b) Feature selection: Select m features that represent the health status of the chiller, which are composed of the chiller's direct monitoring parameters or their simple calculated parameters;

[0055] 1c) Based on the thermodynamic model including mass conservation, momentum conservation, energy conservation and heat and mass transfer mechanism, the thermodynamic change law between fault and feature is analyzed to determine the known fault F i The association rules between the features and the known fault F i Features with strong thermodynamic relationships, where the magnitude of the change in the feature caused by the fault exceeds a pre-set threshold;

[0056] 1d) Determine the structure of the probability graph model: The general structure of the probability graph model constructed by the present invention is shown in the attached Figure 3 , including three layers, from top to bottom, namely, the diagnosis layer, the fault layer and the symptom layer. The diagnosis layer is a discrete node D with n+1 states, representing n known fault states (F1, F2, ..., F n ) and a new fault state (NF); the fault layer includes n fault nodes, each of which represents a specific known fault F i , there are two states, "True" and "False", which represent the known fault Fi Occurrence and non-occurrence; in the symptom layer, each known fault node F i They are connected to three types of nodes, namely, a residual node R, a data node X and several knowledge nodes K. The residual node R and the data node X are m-dimensional continuous nodes composed of m features selected in step 1b). Their data obey the m-dimensional Gaussian distribution. Each knowledge node K represents a feature, which has three states: "Higher", "Normal" and "Lower", which respectively indicate that the measured value of the feature is "significantly increased", "basically unchanged" and "significantly decreased" relative to the baseline value of the feature when the chiller is operating normally. For a known fault node F i , all the known faults F determined in step 1c) i Features with strong thermodynamic relationships are identified as known faults F i Knowledge nodes;

[0057] 1e) Determining the parameters of the probabilistic graphical model comprises the following steps:

[0058] 1e-1) According to expert experience or the frequency of historical failures of the statistical unit, the prior probability of n+1 states of the diagnosis layer node D is determined, as shown in Table 1, where P1, P2, ..., P n , P (n+1) represents the probability value;

[0059] Table 1

[0060]

[0061] 1e-2) Determine the known faults F in the fault layer according to the following three rules: i Conditional probabilities of two states:

[0062] Rule 1: P(F i =True|D=F i )=1,P(F i =False|D=F i )=0;

[0063] Rule 2: P(F i =True|D=F j )=0,P(F i =False|D=F j )=1,i≠j;

[0064] Rule 3: P(F i =True|D=NF)=0,P(F i =False|D=NF)=1; where NF indicates a new fault.

[0065] 1e-3) Determine the conditional probability distribution of the residual node R in the symptom layer, including the following steps:

[0066] 1e-3-1) Selecting a benchmark value model for determining the feature benchmark value, specifically: selecting a regression model with the highest prediction degree from the existing publicly released regression models;

[0067] 1e-3-2) Selecting a suitable operating condition indication feature to indicate the operating condition of the chiller; the suitable operating condition indication feature specifically refers to a feature that can indicate the operating condition of the chiller;

[0068] 1e-3-3) using the operating condition indication feature selected in step 1e-3-2) as an input to the benchmark model;

[0069] 1e-3-4) Using the m features selected in step 1b) as the output of the benchmark model;

[0070] 1e-3-5) Using the normal operating condition data in the training set determined in step 1a-3), the baseline value model is trained to determine the regression coefficient in the baseline value model;

[0071] 1e-3-6) Using the fault operating condition data in the training set determined in step 1a-3), input into the benchmark value model trained in step 1e-3-5), for the m features selected in step 1b), compare the benchmark value output by the model with the actual measured value to obtain the residual data of the two;

[0072] 1e-3-7) By performing maximum likelihood estimation on the residual data obtained in step 1e-3-6), the known fault nodes F in the fault layer are obtained. i The mean vector (μ) describing the conditional probability distribution of the residual node R when it is in the "True" state r，i ) and the covariance matrix (∑ r，i );

[0073] 1e-3-8) Determine the conditional probability distribution of the residual node R in the symptom layer, as shown in Table 2:

[0074] Table 2

[0075] <![CDATA[F i ]]> R True <![CDATA[R~N(μ r，i ,∑ r，i )]]> False <![CDATA[R~N(μ r，i ,c×∑ r，i )]]>

[0076] The coefficient c in Table 2 is determined according to the following two formulas:

[0077]

[0078]

[0079] Where: m represents the dimension of the residual node R, N represents the number of samples, F α(m, Nm) represents the α quantile of the Fisher distribution with degrees of freedom m and Nm, α is the significance level, which is determined by multiple attempts based on the principle of obtaining the best diagnostic performance;

[0080] 1e-4) Determine the conditional probability distribution of the data node X in the symptom layer, as follows: By performing maximum likelihood estimation on the fault operating condition data obtained in step 1a-3), obtain the known fault node F in the fault layer i The mean vector (μ) describing the conditional probability distribution of data node X when it is in the "True" state x，i ) and the covariance matrix (∑ x，i ), and then determine the conditional probability distribution of data node X, as shown in Table 3, where the coefficient c comes from step 1e-3-8);

[0081] Table 3

[0082] <![CDATA[F i ]]> X True <![CDATA[X~N(μ x，i ,∑ x，i )]]> False <![CDATA[X~N(μ x，i ,c×∑ x，i )]]>

[0083] 1e-5) According to expert experience or historical data statistics under fault operation conditions, the conditional probability of the knowledge node K in the symptom layer under three states is determined, as shown in Table 4, where a, b, d, and e represent probability values, and K j represents the jth feature;

[0084] Table 4

[0085]

[0086] 1f) Constructing a benchmark value model of the characteristic benchmark value in the knowledge node K, comprising the following steps:

[0087] 1f-1) Selecting a benchmark value model for determining the feature benchmark value in the knowledge node K, specifically: selecting a regression model with the highest prediction degree from existing publicly released regression models;

[0088] 1f-2) using the operating condition indicator feature selected in step 1e-3-2) as input to the benchmark model;

[0089] 1f-3) Using the features in the knowledge node K selected in step 1d) as the output of the benchmark value model;

[0090] 1f-4) using the normal operating condition data in the training set determined in step 1a-3), training the benchmark value model, determining the regression coefficient in the benchmark value model, and obtaining a trained benchmark value model;

[0091] 1g) Determine the method and judgment threshold for judging whether the feature in the knowledge node K has changed significantly relative to the baseline value, specifically: based on statistical theory, select a mature significance test method, and combine experience to determine the threshold of the selected significance test method.

[0092] Step 2: Online model application, the steps are as follows:

[0093] 2a) Online real-time monitoring of various characteristics of chiller operation;

[0094] 2b) Steady-state filtering: using the same steady-state filtering method as steps 1a-2) to perform steady-state filtering on the data monitored in real time in step 2a) to obtain steady-state data under the real-time operating conditions of the chiller;

[0095] 2c) Feature selection: Select the same features as in step 1b);

[0096] 2d) Obtaining the residual data of the features in the residual node R in the symptom layer, specifically as follows: inputting the real-time steady-state data of the chiller operation obtained in step 2b) into the benchmark value model trained in step 1e-3-5), comparing the benchmark value output by the model with the actual measured value, and obtaining the residual data of the two;

[0097] 2e) obtaining direct monitoring data of the features in the data node X, namely: the steady-state data of the chiller under the real-time operating condition obtained in step 2b);

[0098] 2f) obtaining the state of the feature in the knowledge node K, comprising the following steps: 2f-1) inputting the steady-state data of the chiller under the real-time operating condition obtained in step 2b) into the benchmark value model trained in step 1f-4) to obtain the benchmark value output by the model;

[0099] 2f-2) For the feature in the knowledge node K selected in step 1d), based on the feature change significance judgment method and judgment threshold determined in step 1g), the feature baseline value obtained in step 2f-1) and the actual measured value are compared to obtain the judgment results of "significantly increased", "basically unchanged" and "significantly decreased" of the feature relative to the baseline value;

[0100] 2g) The residual data of the feature in the residual node R obtained in step 2d), the direct monitoring data of the feature in the data node X obtained in step 2e), and the state of the feature in the knowledge node K obtained in step 2f) are input as evidence into the probability graph model constructed in step 1 to obtain the posterior probabilities of the n+1 states of the diagnosis layer node D at this time. According to the maximum a posteriori probability principle, the fault diagnosis result is output, specifically: the state of the diagnosis layer node D with the maximum a posteriori probability is determined as the fault state of the chiller at this time.

[0101] Embodiment 2: This embodiment is a composite fault diagnosis system for chillers based on the joint drive of residual-knowledge-data, including a memory and a processor, wherein the memory stores a computer program for execution on the processor; when the processor executes the computer program, it implements the composite fault diagnosis method for chillers provided in the above-mentioned embodiment 1.

[0102] Embodiment 3: The specific implementation of the residual-knowledge-data joint-driven fault composite diagnosis method provided by the present invention and the beneficial effects of the present invention are further explained below through specific embodiments.

[0103] This embodiment is to diagnose the fault of a centrifugal chiller with a cooling capacity of about 316kW used in the ASHRAE RP-1043 report. The evaporator and condenser of the chiller are both shell and tube heat exchangers, the refrigerant is R134a, and a thermal expansion valve is used. Through the experiment, a large amount of data of 64 parameters of the chiller under normal operating conditions and 7 faults at 4 degradation levels were obtained. The 7 faults are reduced cooling water, reduced chilled water, insufficient refrigerant filling, excessive refrigerant filling, condenser scaling, presence of non-condensable gas and presence of excessive oil.

[0104] Step 1: Probabilistic graphical model construction

[0105] 1a) Data preprocessing, including the following steps:

[0106] 1a-1) Performing a fault test on the chiller to obtain data of the chiller under normal and fault operating conditions;

[0107] 1a-2) Steady-state filtering: using a steady-state filtering method to filter the abnormal and non-steady-state data in the normal and fault data obtained in step 1a-1), and obtain steady-state data of the chiller under normal and fault operating conditions;

[0108] 1a-3) For normal and each fault condition, randomly select 1000 groups of steady-state data obtained in step 1a-2) to form a training set;

[0109] 1b) Feature selection: Select features that characterize the health status of the chiller, which are composed of the direct monitoring parameters of the chiller or their simple calculated parameters. In this embodiment, 14 features are selected, as shown in Table 5, which includes the specific meanings and calculation methods of the selected features, where TCI and TCO represent the cooling water inlet and outlet temperatures, respectively, and TEI and TEO represent the freezing water inlet and outlet temperatures, respectively.

[0110] Table 5

[0111]

[0112] 1c) Based on the conservation of mass, momentum, energy and heat and mass transfer mechanism, the thermodynamic change law between the fault and the feature is analyzed to determine the known fault F i The association rules between the features are shown in Table 6. In the table, “▲” indicates that when a fault occurs, the feature value increases significantly relative to the baseline value. Indicates that when a fault occurs, the relative value of the feature drops significantly, and "-" indicates that when a fault occurs, the relative value of the feature remains basically unchanged. Thus, 7 types of features with strong thermodynamic relationships of known faults are obtained; in Table 6, ε sc It represents the heat exchange rate of the subcooling section, and the calculation formula is as follows:

[0113]

[0114] Table 6

[0115]

[0116] 1d) Determine the structure of the probability graph model: The structure of the probability graph model constructed in this embodiment is shown in the attached Figure 4 , attached Figure 4 Among them, F-1 indicates a cooling water quantity reduction fault, F-2 indicates a chilled water quantity reduction fault, F-3 indicates a refrigerant undercharge fault, F-4 indicates a refrigerant overcharge fault, F-5 indicates a condenser scaling fault, F-6 indicates an excess oil fault, and F-7 indicates a non-condensable gas fault;

[0117] 1e) Determining the parameters of the probabilistic graphical model comprises the following steps:

[0118] 1e-1) Based on expert experience, the prior probabilities of the eight states of the diagnosis layer node D are determined, as shown in Table 7;

[0119] Table 7

[0120]

[0121] 1e-2) Determine the faulty node F in the fault layer i The conditional probabilities of the two states are shown in Table 8. In the table, F-1 indicates a cooling water reduction fault, F-2 indicates a chilled water reduction fault, F-3 indicates a refrigerant undercharge fault, F-4 indicates a refrigerant overcharge fault, F-5 indicates a condenser fouling fault, F-6 indicates an excess oil fault, F-7 indicates a non-condensable gas fault, and NF indicates a new fault;

[0122] Table 8

[0123]

[0124] 1e-3) Determine the conditional probability distribution of the residual node R in the symptom layer, including the following steps:

[0125] 1e-3-1) Selecting a reference value model for determining a feature reference value. In this embodiment, a support vector regression model is selected as the reference value model;

[0126] 1e-3-2) Select appropriate characteristic parameters to indicate the operating condition of the chiller. In this embodiment, the three characteristics of the chilled water outlet temperature, the cooling water inlet temperature and the load rate are selected as the indicating characteristics of the operating condition of the chiller;

[0127] 1e-3-3) using the operating condition indication feature selected in step 1e-3-2) as an input to the benchmark model;

[0128] 1e-3-4) Using the 14 features selected in step 1b) as the output of the benchmark model;

[0129] 1e-3-5) Using the normal operating condition data in the training set determined in step 1a-3), the baseline value model is trained to determine the regression coefficient in the baseline value model;

[0130] 1e-3-6) Use the fault operating condition data in the training set determined in step 1a-3) and input it into the benchmark value model trained in step 1e-3-5), and for the 14 features selected in step 1b), compare the benchmark value output by the model with the actual measured value to obtain the residual data of the two;

[0131] 1e-3-7) By performing maximum likelihood estimation on the residual data obtained in step 1e-3-6), the fault layer node F is obtained. i The mean vector (μ) describing the conditional probability distribution of the residual node R when it is in the "True" state r，i ) and the covariance matrix (∑ r，i );

[0132] 1e-3-8) After multiple trials, this embodiment selects the significance level α=0.025, according to the following formula:

[0133]

[0134]

[0135] The calculation results in c=4, and the conditional probability distribution of the residual node R in the symptom layer is determined, as shown in Table 9:

[0136] Table 9

[0137] <![CDATA[F i ]]> R True <![CDATA[R~N(μ r，i ,∑ r，i )]]> False <![CDATA[R~N(μ r，i ,4×∑ r，i )]]>

[0138] 1e-4) Determine the conditional probability distribution of the data node X in the symptom layer, and obtain the fault layer node F by performing maximum likelihood estimation on the fault operating condition data obtained in step 1a-3). i The mean vector (μ) describing the conditional probability distribution of data node X when it is in the "True" state x，i ) and the covariance matrix (∑ x，i ), and then determine the conditional probability distribution of data node X, as shown in Table 10:

[0139] Table 10

[0140] <![CDATA[F i ]]> X True <![CDATA[X~N(μ x，i ,∑ x，i )]]> False <![CDATA[X~N(μ x，i ,4×∑ x，i )]]>

[0141] 1e-5) Based on expert experience, determine the conditional probabilities of the knowledge node K in the symptom layer under three states, as shown in Tables 11 to 17;

[0142] Table 11

[0143]

[0144] Table 12

[0145]

[0146] Table 13

[0147]

[0148] Table 14

[0149]

[0150] Table 15

[0151]

[0152] Table 16

[0153]

[0154] Table 17

[0155]

[0156] 1f) Constructing a benchmark value model of the characteristic benchmark value in the knowledge node K, comprising the following steps:

[0157] 1f-1) Selecting a reference value model for determining the feature reference value in the knowledge node K. In this embodiment, a support vector regression model is selected as the reference value model;

[0158] 1f-2) using the operating condition indicator feature selected in step 1e-3-2) as input to the benchmark model;

[0159] 1f-3) Using the features in the knowledge node K selected in step 1d) as the output of the benchmark value model;

[0160] 1f-4) using the normal operating condition data in the training set determined in step 1a-3), training the benchmark value model, determining the regression coefficient in the benchmark value model, and obtaining a trained benchmark value model;

[0161] 1g) This embodiment selects the t-statistic method in statistics as a method for judging whether the change in the reference value is significant, and determines the threshold value to be ±3σ, where σ is the standard deviation of the characteristic reference value data;

[0162] Step 2: Application of online model

[0163] 2a) Online real-time monitoring of various characteristics of chiller operation;

[0164] 2b) Steady-state filtering: using the same steady-state filtering method as steps 1a-2) to perform steady-state filtering on the data monitored in real time in step 2a) to obtain steady-state data under the real-time operating conditions of the chiller;

[0165] In order to evaluate the performance of the residual-knowledge-data jointly driven chiller fault composite diagnosis method provided in this embodiment, 400 groups of steady-state data are randomly selected from the experimental data after steady-state filtering in the above steps 1a-2) for 7 types of faults including all 4 degradation levels to form a test set to test the fault diagnosis performance of the constructed residual-knowledge-data jointly driven chiller fault composite diagnosis model;

[0166] 2c) Feature selection: Select the same features as in step 1b);

[0167] 2d) Obtaining the residual data of the features in the residual node R in the symptom layer, specifically as follows: inputting the real-time steady-state data of the chiller operation obtained in step 2b) into the benchmark value model trained in step 1e-3-5), comparing the benchmark value output by the model with the actual measured value, and obtaining the residual data of the two;

[0168] 2e) obtaining direct monitoring data of the features in the data node X, namely: the steady-state data of the chiller under the real-time operating condition obtained in step 2b);

[0169] 2f) Obtaining the state of the feature in the knowledge node K, comprising the following steps:

[0170] 2f-1) inputting the steady-state data of the chiller under real-time operating conditions obtained in step 2b) into the benchmark value model trained in step 1f-4) to obtain the benchmark value output by the model;

[0171] 2f-2) For the feature in the knowledge node K selected in step 1d), based on the feature change significance judgment method and judgment threshold determined in step 1g), the feature baseline value obtained in step 2f-1) and the actual measured value are compared to obtain the judgment results of "significantly increased", "basically unchanged" and "significantly decreased" of the feature relative to the baseline value;

[0172] 2g) The residual data of the feature in the residual node R obtained in step 2d), the direct monitoring data of the feature in the data node X obtained in step 2e), and the state of the feature in the knowledge node K obtained in step 2f) are input as evidence into the probability graph model constructed in step 1 to obtain the posterior probabilities of the 8 states of the diagnosis layer node D at this time. According to the maximum a posteriori probability principle, the fault diagnosis result is output, specifically: the state of the diagnosis layer node D with the maximum a posteriori probability is determined as the fault state of the chiller at this time.

[0173] Diagnostic performance evaluation: Use all the test set data determined in step 2b) of this embodiment to perform performance evaluation on the fault diagnosis method provided in this embodiment.

[0174] The performance evaluation results are shown in Table 18, where the first column represents 7 known fault categories, the second column represents 4 degradation levels, from low to high, namely SL-1, SL-2, SL-3 and SL-4, the third column X represents the diagnostic accuracy using only the data node evidence from the symptom layer, the fourth column R represents the diagnostic accuracy using only the residual node evidence from the symptom layer, the fifth column K represents the diagnostic accuracy using only the knowledge node evidence from the symptom layer, and the sixth column X+R+K represents the diagnostic accuracy using all the node evidence from the symptom layer at the same time. From Table 18, compared with the diagnostic results using only the residual node or data node or knowledge node evidence from the symptom layer, the diagnostic accuracy when using the three types of node evidence at the same time is significantly higher, especially for faults at low degradation levels (SL-1 and SL-2), the maximum improvement in diagnostic accuracy is about 43% (for the refrigerant undercharge fault under SL-1, the three types of node evidence are used simultaneously compared to the use of only knowledge node evidence).

[0175] Table 18

[0176]

[0177] The results of this embodiment fully show that the residual-knowledge-data jointly driven composite diagnosis method for chiller faults provided by the present invention successfully integrates the advantages of the three types of models, namely, residual-driven, knowledge-driven and data-driven, into the same diagnostic system, achieving complementary advantages and obtaining better diagnostic performance. Specifically, the provided residual-knowledge-data jointly driven composite diagnosis method for chiller faults can tolerate the absence of any type of node in the diagnostic process, and can use any type of node alone to complete the fault diagnosis task; the provided composite diagnosis method fully inherits the advantages of the three types of methods, namely, residual, knowledge and data, and makes up for their disadvantages to a certain extent, and can be applied to sites where the three types of nodes are not applicable, and when the three types of nodes are used simultaneously, the optimal diagnostic performance is obtained.

[0178] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Those skilled in the art should understand that the specific implementation modes of the present invention may be modified or replaced by equivalents with reference to the above embodiments. Any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention are within the scope of protection of the pending claims.

Claims

1. A residual-knowledge-data joint-driven fault composite diagnosis method for chiller fault diagnosis, characterized in that: The steps include: (1) obtaining data of the chiller under normal and faulty operating conditions, filtering the abnormal and non-steady-state data in the normal and faulty data in a steady state, obtaining the steady-state data of the chiller under normal and faulty operating conditions, and selecting an indicator feature for determining the operating condition of the chiller; (2) Determine the structure of the probabilistic graphical model; the structure of the probabilistic graphical model includes three layers, from top to bottom, namely, the diagnosis layer, the fault layer, and the symptom layer, where the diagnosis layer is a discrete node D ,have n +1 status, representing n known fault states and 1 new fault state; the fault layer includes n Fault nodes, each fault node represents a specific known fault F i , with two states of "True" and "False", indicating known faults F i happens and doesn’t happen; In the symptom layer, each known fault F i They are connected to three types of nodes, namely 1 residual node R , 1 data node X and several knowledge nodes K , residual node R and data nodes X The health status of the chiller is characterized by m Features m Dimensional continuous nodes, whose data obey m Gaussian distribution, each knowledge node K Each represents a feature, and there are three states, which respectively represent that the measured value of the feature is "significantly increased", "basically unchanged" and "significantly decreased" relative to the baseline value of the feature when the chiller is operating normally; Knowledge node K The specific type is determined by the association rules between faults and features; (3) Determining the parameters of the probability graph model; the parameters of the probability graph model include the diagnosis layer nodes D of n +1 prior probability of state, known faults in the fault layer F i Conditional probabilities of two states, residual nodes in the symptom layer R and data nodes X The conditional probability distribution of the knowledge nodes in the symptom layer K The conditional probability of the diagnosis layer node D of n+ The prior probability of a state is obtained from expert experience or the frequency of historical failures of the statistical unit; the known faults in the fault layer F i The conditional probabilities of the two states are assigned according to the following three rules: Rule 1: ; Rule 2: ; Rule 3: ;in NF Indicates a new fault; The residual nodes in the symptom layer R The conditional probability distribution of the known faults F i When the residual data of each feature is in the "True" state, it is obtained by maximum likelihood estimation; the data nodes in the symptom layer X The conditional probability distribution of the known faults F i When the state is "True", the direct monitoring steady-state data of each feature is obtained by maximum likelihood estimation; the knowledge nodes in the symptom layer K The conditional probabilities under the three states are obtained from expert experience or historical data statistics under faulty operating conditions; (4) Select the node to determine the knowledge K The benchmark value model of the characteristic benchmark value will constitute the knowledge node K The characteristics of are used as the output of the model, and the indicative characteristics of the chiller operating conditions are used as the input of the model. The steady-state data of the chiller under normal operating conditions are used to train the benchmark value model to obtain a constructed benchmark value model. (5) Select a method to judge the significance of feature changes and K Whether the characteristic has changed significantly relative to the reference value is judged, and the judgment threshold of the characteristic change significance judgment method is determined; (6) Obtain the real-time data of the chiller during operation, and after steady-state filtering and feature selection, obtain the residual nodes in the symptom layer R Residual data and data nodes of features in X Direct monitoring data of features and knowledge nodes K The state of the features in the model is used as evidence and input into the constructed probability graph model to calculate the diagnosis layer node D The posterior probability of each state is used to determine what failure the chiller has.

2. The residual-knowledge-data joint-driven fault composite diagnosis method according to claim 1 is characterized by: In step (2), for a known fault F i , knowledge node K The specific type is determined by the association rules between faults and features. The specific method is as follows: first, based on the thermodynamic principles, the thermodynamic change laws between faults and features are analyzed to determine the known faults. F i and the association rules between the features; then all F i Features with strong thermodynamic relationships are identified as known faults F i The connected knowledge nodes, the strong thermodynamic relationship is that the magnitude of the change in the fault-induced characteristic exceeds a preset threshold.

3. The residual-knowledge-data jointly driven fault composite diagnosis method according to claim 1 is characterized in that: Step (3) is to identify the known fault F i When it is the residual data of the feature in the "True" state, it includes the following steps: (a) Select the reference value model used to determine the feature reference value, which will form the residual node R The characteristics of are taken as the output of the model, and the indicative characteristics of the chiller operating conditions are taken as the input of the model; (b) using the acquired steady-state data of the chiller under normal operating conditions to train the benchmark value model, and obtain a trained characteristic benchmark value model; (c) Use the acquired steady-state data of the chiller under faulty operating conditions and input it into the trained feature benchmark value model. For each feature, compare the benchmark value output by the model with the actual measured value to obtain the residual data of the two, which is the residual data of the feature.

4. The residual-knowledge-data joint-driven fault composite diagnosis method according to claim 1 is characterized in that: In step (6), the residual node in the symptom layer is obtained R The method for obtaining the residual data of the features in the data node is as follows: input the obtained real-time data of the chiller operation into the trained feature benchmark value model, and for each feature, compare the benchmark value output by the model with the actual measured value to obtain the residual data of the two; data node X The direct monitoring data of the feature is the real-time data of the chiller operation; the knowledge node is obtained K The method for determining the state of the features in the model is as follows: the real-time data of the chiller operation is input into the trained feature benchmark value model to obtain the benchmark value output by the model. For each feature, based on the feature change significance judgment method and the determined judgment threshold, according to the benchmark value output by the feature benchmark value model and the actual measured value, the judgment results of "significantly increased", "basically unchanged" and "significantly decreased" of the feature relative to the benchmark value are obtained, thereby forming a knowledge node. K The status of the feature.

5. A residual-knowledge-data jointly driven fault composite diagnosis system, comprising a memory and a processor, wherein the memory stores a computer program for execution on the processor; characterized in that: When the processor executes the computer program, the following steps are implemented: (1) obtaining data of the chiller under normal and faulty operating conditions, filtering the abnormal and non-steady-state data in the normal and faulty data in a steady state, obtaining the steady-state data of the chiller under normal and faulty operating conditions, and selecting an indicator feature for determining the operating condition of the chiller; (2) Determine the structure of the probabilistic graphical model; the structure of the probabilistic graphical model includes three layers, from top to bottom, namely, the diagnosis layer, the fault layer, and the symptom layer, where the diagnosis layer is a discrete node D ,have n +1 status, representing n known fault states and 1 new fault state; the fault layer includes n Fault nodes, each fault node represents a specific known fault F i , with two states of "True" and "False", indicating known faults F i happens and doesn’t happen; In the symptom layer, each known fault F i They are connected to three types of nodes, namely 1 residual node R , 1 data node X and several knowledge nodes K , residual node R and data nodes X The health status of the chiller is characterized by m Features m Dimensional continuous nodes, whose data obey m Gaussian distribution, each knowledge node K Each represents a feature, and there are three states, which respectively represent that the measured value of the feature is "significantly increased", "basically unchanged" and "significantly decreased" relative to the baseline value of the feature when the chiller is operating normally; Knowledge node K The specific type is determined by the association rules between faults and features; (3) Determining the parameters of the probability graph model; the parameters of the probability graph model include the diagnosis layer nodes D of n +1 prior probability of state, known faults in the fault layer F i Conditional probabilities of two states, residual nodes in the symptom layer R and data nodes X The conditional probability distribution of the knowledge nodes in the symptom layer K The conditional probability of the diagnosis layer node D of n+ The prior probability of a state is obtained from expert experience or the frequency of historical failures of the statistical unit; the known faults in the fault layer F i The conditional probabilities of the two states are assigned according to the following three rules: Rule 1: ; Rule 2: ; Rule 3: ;in NF Indicates a new fault; The residual nodes in the symptom layer R The conditional probability distribution of the known faults F i When the residual data of each feature is in the "True" state, it is obtained by maximum likelihood estimation; the data nodes in the symptom layer X The conditional probability distribution of the known faults F i When the state is "True", the direct monitoring steady-state data of each feature is obtained by maximum likelihood estimation; the knowledge nodes in the symptom layer K The conditional probabilities under the three states are obtained from expert experience or historical data statistics under faulty operating conditions; (4) Select the node to determine the knowledge K The benchmark value model of the characteristic benchmark value will constitute the knowledge node K The characteristics of are used as the output of the model, and the indicative characteristics of the chiller operating conditions are used as the input of the model. The steady-state data of the chiller under normal operating conditions are used to train the benchmark value model to obtain a constructed benchmark value model. (5) Select a method to judge the significance of feature changes and K Whether the characteristic has changed significantly relative to the reference value is judged, and the judgment threshold of the characteristic change significance judgment method is determined; (6) Obtain the real-time data of the chiller during operation, and after steady-state filtering and feature selection, obtain the residual nodes in the symptom layer R Residual data and data nodes of features in X Direct monitoring data of features and knowledge nodes K The state of the features in the model is used as evidence and input into the constructed probability graph model to calculate the diagnosis layer node D The posterior probability of each state is used to determine what failure the chiller has.

6. The residual-knowledge-data jointly driven fault composite diagnosis system according to claim 5, characterized in that: In step (2), for a known fault F i , knowledge node K The specific type is determined by the association rules between faults and features. The specific method is as follows: first, based on the thermodynamic principles, the thermodynamic change laws between faults and features are analyzed to determine the known faults. F i and the association rules between the features; then all F i Features with strong thermodynamic relationships are identified as known faults F i The connected knowledge nodes, the strong thermodynamic relationship is that the magnitude of the change in the fault-induced characteristic exceeds a preset threshold.

7. The residual-knowledge-data jointly driven fault composite diagnosis system according to claim 5, characterized in that: Step (3) is to identify the known fault F i When it is the residual data of the feature in the "True" state, it includes the following steps: (a) Select the reference value model used to determine the feature reference value, which will form the residual node R The characteristics of are taken as the output of the model, and the indicative characteristics of the chiller operating conditions are taken as the input of the model; (b) using the acquired steady-state data of the chiller under normal operating conditions to train the benchmark value model, and obtain a trained characteristic benchmark value model; (c) Use the acquired steady-state data of the chiller under faulty operating conditions and input it into the trained feature benchmark value model. For each feature, compare the benchmark value output by the model with the actual measured value to obtain the residual data of the two, which is the residual data of the feature.

8. The residual-knowledge-data jointly driven fault composite diagnosis system according to claim 5, characterized in that: In step (6), the residual node in the symptom layer is obtained R The method for obtaining the residual data of the features in the data node is as follows: input the obtained real-time data of the chiller operation into the trained feature benchmark value model, and for each feature, compare the benchmark value output by the model with the actual measured value to obtain the residual data of the two; data node X The direct monitoring data of the feature is the real-time data of the chiller operation; the knowledge node is obtained K The method for determining the state of the features in the model is as follows: the real-time data of the chiller operation is input into the trained feature benchmark value model to obtain the benchmark value output by the model. For each feature, based on the feature change significance judgment method and the determined judgment threshold, according to the benchmark value output by the feature benchmark value model and the actual measured value, the judgment results of "significantly increased", "basically unchanged" and "significantly decreased" of the feature relative to the benchmark value are obtained, thereby forming a knowledge node. K The status of the feature.

Images