A Fault Diagnosis Method for Closed-Loop Control Systems Based on CIOHMM

Abstract

The present invention relates to a fault diagnosis method for a closed-loop control system based on CIOHMM. A feedback neural network and a Kalman filter / particle filter method are used to extract fault features, thereby obtaining a feature data set of a main fault type and establishing a CIOHMM model for realizing target equipment state quantity judgment. Different from traditional HMM and IOHMM which have inputs, a continuous closed-loop control signal of the equipment is set as the input of the CIOHMM model. Different state transfer matrices are used to describe the dynamic change of the state for different inputs, so that the state evolution of the equipment under closed-loop control can be described. The CIOHMM model established in this way not only describes the relationship between the equipment state and the observed value, but also describes the influence of the closed-loop control on the dynamic change of the system state, so that accurate diagnosis of the equipment fault of the air-conditioning unit can be realized, which has unique advantages over other methods.

Description

Technical Field

[0001] The invention relates to a fault diagnosis method for a closed-loop control system based on CIOHMM, and belongs to the technical field of air-conditioning unit fault detection. Background Art

[0002] HVAC systems account for the largest share of total building energy consumption, about 50-60%. About 42% of cooling energy consumption and 26% of maintenance costs in HVAC are caused by equipment failure. It is estimated that HVAC fault diagnosis can reduce energy consumption by 10-40%. As an important subsystem of the HVAC system, the cooling coil is used to reduce the indoor air temperature to its set value. Due to the introduction of closed-loop control, the functional relationship between different variables will change. Once an abnormality occurs in the closed-loop system, since feedback control improves the robustness of the system, in the early stage of the fault, the residual signal is very small, and its small fluctuation will be suppressed by feedback control, thus greatly reducing the success rate of fault detection. In addition, since the fault will propagate through feedback control, it will eventually cause multiple devices to have abnormalities, making it difficult to locate and isolate the fault. Based on the above two reasons, many mature fault diagnosis methods used in open-loop systems are no longer applicable, so it is very important to diagnose the fault of cooling coils under closed-loop control. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a fault diagnosis method for a closed-loop control system based on CIOHMM, which adopts a new design strategy and can realize efficient diagnosis of air-conditioning unit faults.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions: the present invention designs a fault diagnosis method for a closed-loop control system based on CIOHMM, and executes the following steps A to C for each target device of the air-conditioning unit to obtain a fault diagnosis model corresponding to the target device; and then according to step i, applies the fault diagnosis model corresponding to the target device to realize fault diagnosis of the target device;

[0005] Step A. Obtain each target fault feature corresponding to the target device, and obtain the data value of each target fault feature corresponding to each preset historical time node, construct the target fault feature data matrix corresponding to the target device, and then enter step B;

[0006] Step B. Construct the initial probability vector π corresponding to the target device = {π i}、State transfer matrix P={a ij (U)}, observation likelihood function b i(O), considering the target device control quantity and the target device corresponding to each target fault characteristic observation value, a CIOHMM model for judging the target device state quantity is implemented, wherein the target device control quantity is used to represent the working mode of the target device, and the target device state quantity is used to represent that the target device is in a normal working state or in various fault states, i = 1, ..., N, j = 1, ..., N, N represents the number of states of the target device, π i represents the initial probability of the target device being in the i-th state, a ij (U) represents the probability of transition from the i-th state to the j-th state under the target device control quantity U, O represents the target device corresponding to each target fault characteristic observation value, b i (O) represents the probability of each target fault feature observation value of the target device corresponding to the i-th state; then proceed to step C;

[0007] Step C. Based on the target fault feature data matrix corresponding to the target device, combined with the target fault features, the target device control quantity and the target device state quantity corresponding to each preset historical time node are respectively corresponded, and the EM algorithm is used to train the CIOHMM model corresponding to the target device to obtain the trained CIOHMM model, that is, the fault diagnosis model corresponding to the target device;

[0008] Step i. Obtain the observed value of each target fault feature of the target device at the target time point, combine it with the control value of the target device at the target time point, apply the fault diagnosis model corresponding to the target device, obtain the state value of the target device at the target time point through the Viterbi algorithm, and realize the fault diagnosis of the target device at the target time point.

[0009] As a preferred technical solution of the present invention: the step A includes the following steps A1 to A4;

[0010] Step A1. Define the difference between the data of each variable with a set value involved in the target device and its set value, and preset each sensor detection related to the fault involved in the target device as each fault feature to be analyzed, and proceed to step A2;

[0011] Step A2. Based on the feedback neural network model corresponding to the pre-trained target device, which takes the preset target detection feature data as input and the preset target analysis feature estimation value as output, define the difference between the target analysis feature estimation value output by the feedback neural network model and the target analysis feature calculation value as the fault feature to be analyzed, and then enter step A3;

[0012] Step A3. Use the Kalman filter / particle filter method to estimate and obtain various parameters related to the fault involved in the target device as the fault characteristics to be analyzed, and then proceed to step A4;

[0013] Step A4. Based on the data values ​​of each preset historical time node corresponding to each fault feature to be analyzed, the PCA principal component analysis method is used to obtain the main component fault features to be analyzed in each fault feature to be analyzed as the target fault features corresponding to the target device, and the data values ​​of each target fault feature corresponding to each preset historical time node are obtained to construct the target fault feature data matrix corresponding to the target device.

[0014] As a preferred technical solution of the present invention: if the target device is a cooling coil, in step A2, based on the pre-trained cooling coil corresponding to the chilled water flow Mixed air flow Mixed air temperature T a,mix , Chilled water supply temperature T chw,sup is the input and heat transfer Q cc The estimated value is the feedback neural network model with the output, combined with Heat exchange Q cc Calculated value, defining the heat transfer Q output by the feedback neural network model cc Estimated value and heat transfer Q cc The difference between the calculated values ​​is used as the fault feature to be analyzed, where c pw represents the specific heat capacity of chilled water, T chw,rn Indicates the chilled water return temperature.

[0015] As a preferred technical solution of the present invention: the step C includes the following steps C1 to C4;

[0016] Step C1. About π={π i}, according to the following formula:

[0017]

[0018] Calculate the initial probability vector π = {π i}, where I 1 Indicates the state of the target device at the first moment, I 1 =i means that the state of the target device corresponding to the first moment is the i-th state, O means the target device corresponds to the target fault feature observation value of all moments, λ = {π,P,b i (O 1 )},b i (O 1 ) represents the probability of each target fault characteristic observation value of the target device at the first moment corresponding to the i-th state; U represents the control amount of the target device corresponding to all moments, and Pr(·) represents the probability function;

[0019] Step C2. Regarding the state transfer matrix P = {a ij (U)}, first obtain the following formula:

[0020]

[0021] 1≤k≤K, K represents the total number of moments, I k Indicates the state of the target device at the kth moment, U k represents the control amount of the target device at the kth moment, a 0,ij 、a 1,ij 、a 2,ij represents the constant term, U k Linear coefficient, U k The coefficient of the quadratic term;

[0022] Then derive the following formula:

[0023]

[0024] Among them, a IkIk+1 (U k ) represents the probability of the target device transferring from the state at the kth moment to the state at the k+1th moment under the control quantity at the kth moment, and I represents the state of the target device corresponding to all moments;

[0025] Combination Right now Based on the Lagrange multiplier method, the following function is obtained:

[0026]

[0027] in, represents the Lagrange multiplier, and combined with Then using the chain method, we get:

[0028]

[0029] in,

[0030] First, a ij (U k ) Take the partial derivative and get the following:

[0031]

[0032] Then we further obtain the following:

[0033]

[0034] Summing j on both sides of the equal sign, we get the following:

[0035]

[0036] Substituting the above formula into Obtained as follows:

[0037]

[0038]

[0039] Then we further deduce a 1,ij The iteration formula is as follows:

[0040]

[0041] The same goes for a 2,ij Taking the derivative, we can also get

[0042]

[0043] By using the least squares method, we can estimate a from the above formula 0,ij ,a 1,ij and a 2,ij ,i,j=1,…,N to obtain the state transfer matrix P;

[0044] Step C3. The target fault feature observation values ​​O corresponding to the i-th state of the target device obey the mixed Gaussian distribution as follows:

[0045]

[0046] in:

[0047]

[0048] C represents the number of Gaussian models in the mixed Gaussian distribution, c ic represents the weight of the cth Gaussian model of the mixed normal distribution obeyed by each target fault feature observation value corresponding to the i-th state of the target device, f(O|μ ic ,U ic ) is the probability density function of the cth Gaussian model in the mixed Gaussian distribution corresponding to each target fault feature observation value in the ith state, then the distribution parameter estimate of each target fault feature observation value corresponding to the cth Gaussian model of the target device corresponding to the ith state is The calculation formula is as follows:

[0049]

[0050]

[0051] in:

[0052]

[0053] Step C4. Based on the target fault feature data matrix corresponding to the target device, the target device control quantity and the target device state quantity corresponding to each preset historical time node are respectively matched with the target fault feature. Based on the above iterative formula, each parameter is iteratively updated in turn until the preset convergence condition is met, and the initial probability vector π={π i}、State transfer matrix P={a ij (U)}, observation likelihood function b i (O) Obtain the trained CIOHMM model, that is, the fault diagnosis model corresponding to the target device.

[0054] The fault diagnosis method of a closed-loop control system based on CIOHMM described in the present invention adopts the above technical solution and has the following technical effects compared with the prior art:

[0055] The present invention designs a fault diagnosis method for a closed-loop control system based on CIOHMM, which adopts feedback neural network, Kalman filter / particle filter and other methods to extract fault features, thereby obtaining feature data sets of main fault types, and establishing a CIOHMM model for realizing target equipment state quantity judgment; different from traditional HMM and IOHMM with input, the continuous closed-loop control signal of the equipment is set as the input of the CIOHMM model, and different state transfer matrices are used to describe the dynamic change of the state for different inputs, which can describe the state evolution of the equipment under closed-loop control; the CIOHMM model established in this way not only describes the relationship between the equipment state and the observed value, but also describes the influence of closed-loop control on the dynamic change of the system state, which can realize accurate diagnosis of air-conditioning unit equipment faults, and has unique advantages over other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a block diagram of a method for diagnosing faults of an air conditioner unit according to the present invention;

[0057] Figure 2 It is the CIOHMM model framework that describes the health status changes of the air supply fan. DETAILED DESCRIPTION

[0058] The specific implementation modes of the present invention will be further described in detail below in conjunction with the accompanying drawings.

[0059] The present invention designs a closed-loop control system fault diagnosis method based on CIOHMM, targeting each target device of the air-conditioning unit, such as Figure 1 As shown, execute the following steps A to C to obtain the fault diagnosis model corresponding to the target device.

[0060] For air conditioning units, the sensor data involved in fault diagnosis includes the flow rate of chilled water. Chilled water supply temperature T chw,sup , Chilled water return temperature T chw,rn , mixed air flow Air flow Mixed air temperature T a,mix , air supply temperature T a,sup , Power of air supply fan Q sf ; and certain equipment parameters, including the pressure drop ΔP of the supply fan sf,des , supply air temperature setting value T sup,spt , Design value of air supply volume

[0061] Step A. Obtain each target fault feature corresponding to the target device, and obtain the data value of each target fault feature corresponding to each preset historical time node, construct the target fault feature data matrix corresponding to the target device, and then enter step B.

[0062] In actual application, the above step A is specifically designed to execute the following steps A1 to A4.

[0063] Step A1. Define the differences between the data of each variable with set values ​​involved in the target device and its set values, and preset each sensor detection related to the fault involved in the target device as each fault feature to be analyzed, and proceed to step A2.

[0064] If the target device is a cooling coil, then step A1 involves the flow rate of chilled water. Air flow And the supply air temperature T a,sup With its set value T sup,spt The difference between them is taken as the fault feature to be analyzed.

[0065] Step A2. Based on the feedback neural network model corresponding to the pre-trained target device, which takes the preset target detection feature data as input and the preset target analysis feature estimation value as output, define the difference between the target analysis feature estimation value output by the feedback neural network model and the target analysis feature calculation value as the fault feature to be analyzed, and then enter step A3.

[0066] In actual application, if the target device is a cooling coil, then in step A2, based on the pre-trained cooling coil corresponding to the chilled water flow Mixed air flow Mixed air temperature T a,mix , Chilled water supply temperature T chw,sup is the input and heat transfer Q cc The estimated value is the feedback neural network model with the output, combined with Heat exchange Q cc Calculated value, defining the heat transfer Q output by the feedback neural network modelcc Estimated value and heat transfer Q cc The difference between the calculated values ​​is used as the fault feature to be analyzed, where c pw represents the specific heat capacity of chilled water, T chw,rn Indicates the chilled water return temperature.

[0067] Step A3: Use the Kalman filter / particle filter method to estimate and obtain various parameters related to the fault involved in the target device as the fault characteristics to be analyzed, and then enter step A4.

[0068] In the actual application of step A3, the inner diameter d of the pipe wall including the cooling coil is obtained. tube,in and the fin surface area A fin And the efficiency of the air supply fan sf A filtering framework is built based on the gray box model containing these parameters, and the Kalman filter / particle filter method is used to estimate these parameters. For the two key devices in the air conditioning unit, the cooling coil and the fan, different gray box models and different filtering methods are used to estimate the fault-related parameters. A gray box model of the cooling coil contains d tube,in and A fin The two parameters are as follows:

[0069]

[0070] Where, LMTD=(ΔT sup -ΔT rn ) / (lnΔT sup -lnΔT rn ),

[0071] τ=(A fin +A tube,out ) / A tube,in ,

[0072] α a,e =α a [1-A fin (1-η fin ) / (A tube,out +A fin )],

[0073]

[0074] Among them, E a,mix and E a,sup is the enthalpy of the mixed air and the cooled supply air, δ tube is the thickness of the pipe, λ tube is the thermal conductivity of the pipe, R f is the thermal resistance caused by dust, A tube,in is the internal surface area of ​​the pipe, A tube,outis the outer surface area of ​​the pipe, A fin is the outer surface area of ​​the coil fins. ΔT sup =T a,sup –T chw,sup ,ΔT rn =T a,mix –T chw,rn , where T a,sup is the supply air temperature, T chw,rn is the return temperature of chilled water, α a is the convective heat transfer coefficient of the fin, η fin is the fin efficiency, A is the physical coefficient, is the flow rate of chilled water, ψ is the heat flux. tube,in and A fin These two parameters are used as the state x in the filtering framework cc , these parameters are estimated based on the measured values. Since the model is nonlinear, the particle filter framework established for the cooling coil is as follows:

[0075] Equation of state: x cc (t+1)=x cc (t)+v cc (t), where is the state of the cooling coil at time t+1, that is, the inner diameter d of the pipe at time t+1 tube,in (t+1) and fin area A fin The value of (t+1); is the process noise at time t, v tube,in (t) is the process noise of the pipe inner diameter at time t, v fin (t) is the process noise of the fin area at time t, and the state equation describes the dynamic change process of the cooling coil state.

[0076] Measurement equation: z cc (t) = h(x cc (t))+w cc (t), where w cc (t) is the measurement noise at time t. The measurement equation describes the relationship between the measurable variable (sensor reading) and the unmeasurable fault characteristic parameter (d tube,in , A fin ) between them.

[0077] Based on the state equation and the measurement equation, the particle filter method is used to represent the posterior probability of random events through a set of weighted random samples, namely particles, and the state of the cooling coil at each time t is estimated from the sensor readings containing noise in the formula. This is the estimated value of the pipe inner diameter. and an estimate of the fin area

[0078] Based on the fault characteristics of the cooling coil obtained above, a fault feature matrix can be established. The feature matrix X cc As shown below:

[0079]

[0080] A gray box model of the air supply fan includes the fan efficiency e sf This parameter is as follows:

[0081]

[0082] Where Q sf is the power of the air supply fan; e sf is the efficiency of the air supply fan; ρ air is the density of air; is the air supply volume; is the design value of the air supply volume; ΔP sf,des The pressure drop of the supply fan. sf The linear equation of , so Kalman filtering is used to estimate e sf The Kalman filter framework is as follows:

[0083] Equation of state: x sf (t+1)=x sf (t)+v sf (t), where x sf (t+1)=e sf (t+1) is the state of the air supply fan at time t+1; v sf (t) is the process noise at time t. The state equation reflects the dynamic change process of the air supply fan state.

[0084] Measurement equation:

[0085]

[0086] in w sf (t) is the measurement noise.

[0087] The measurement equation describes the relationship between the measurable variable (sensor reading) and the unmeasurable fault characteristic parameter (e sf ). The gray box model is about the fault-related parameters e sf The linear model is used, so the KF method is used to estimate e sf .

[0088] Based on the fault characteristics of the air supply fan, the fault characteristic matrix X of the air supply fan can be established: sf As shown below:

[0089]

[0090] Where K is the length of the time series; ΔT sup,spt =T a,sup –T sup,spt .

[0091] Step A4. Based on the data values ​​of each preset historical time node corresponding to each fault feature to be analyzed, the PCA principal component analysis method is used to obtain the main component fault features to be analyzed in each fault feature to be analyzed as the target fault features corresponding to the target device, and the data values ​​of each target fault feature corresponding to each preset historical time node are obtained to construct the target fault feature data matrix corresponding to the target device.

[0092] In the application, dimensionality reduction processing is to extract the main components of the fault feature matrix to be analyzed to form a new low-dimensional matrix. The specific steps are as follows:

[0093] Assume the feature matrix is:

[0094]

[0095] Where M is the number of feature variables in the feature matrix, and K is the length of the time series. First, a decentralized operation is performed on each row of the fault feature matrix X to be analyzed. For example, the i-th row X i =[x 1 (i) x 2 (i) … x M (i)]

[0096]

[0097] Next, the covariance matrix of the samples in the matrix is ​​calculated.

[0098] C=XX T ,

[0099] Then, find the eigenvalues ​​and corresponding eigenvectors of the covariance matrix C. Arrange the eigenvalues ​​in order of size, and take the eigenvectors corresponding to the first k eigenvalues ​​to form the matrix P. These k eigenvalues ​​as principal components need to contain more than 90% of the information of all components. For the cooling coil, the matrix O composed of the first three principal components is cc can cover 98.4% of the information, so O cc The first two principal components of the fan are used as the observations of the dynamic hidden Markov model of the cooling coil. sf It can cover 96.7% of the information and thus serve as the observation value of the dynamic hidden Markov model of the air supply fan.

[0100] Step B. Construct the initial probability vector π corresponding to the target device = {π i}、State transfer matrix P={a ij (U)}, observation likelihood function b i (O), considering the target device control quantity and the target device corresponding to each target fault characteristic observation value, a CIOHMM model for judging the target device state quantity is implemented, wherein the target device control quantity is used to represent the working mode of the target device, and the target device state quantity is used to represent that the target device is in a normal working state or in various fault states, i = 1, ..., N, j = 1, ..., N, N represents the number of states of the target device, π i represents the initial probability of the target device being in the i-th state, a ij (U) represents the probability of transition from the i-th state to the j-th state under the target device control quantity U, O represents the target device corresponding to each target fault characteristic observation value, b i (O) represents the probability of each target fault feature observation value of the target device corresponding to the i-th state; then enter step C.

[0101] Here, the CIOHMM model that considers the target device control quantity and the target device corresponding to each target fault characteristic observation value is used to realize the target device state quantity judgment. For example, for the air supply fan, the CIOHMM model architecture of its corresponding health state change is as follows: Figure 2 shown.

[0102] In actual applications, regarding the target device control quantity here, the AHU where the cooling coil is located has three control modes, namely 1) heating; 2) cooling with outdoor air; 3) mechanical cooling with 100% outdoor air introduction; and 4) mechanical cooling with minimum outdoor air flow. Different control strategies will be designed in different modes. Therefore, the consideration of the CIOHMM model for realizing the judgment of the target device state quantity here involves the working mode of the target device, that is, the target device control quantity is used to represent the working mode of the target device. Specifically, the control signal CHWC-VLV of the valve opening of the cooling coil is used as the input U of its IOHMM. cc (t) = CHWC-VLV(t). For the air supply fan, there are two main control signals, including 1) the fan start and stop control signal SF-SST; and 2) the fan speed proportional control signal SF-SPD. Therefore, these two variables are used as the input of the fan's IOHMM

[0103] In practical applications, regarding the target device state quantity here, such as the state of the IOHMM model of the cooling coil, the state is a combination of different faults. Consider four types of faults of the cooling coil, including 1) deposition f in the cooling coil pipe tube,cc; 2) Dust accumulates on the fins of the cooling coil fin,cc ; 3) The chilled water valve of the cooling coil is stuck in the fully open state f vlvso,cc ; and 4) the chilled water valve of the cooling coil is stuck in the fully closed state f vlvsc,cc . Then the established IOHMM model has N cc =2 4 = 16 states and is defined as s cc =0,1,…,15, where s cc =0 corresponds to "0000", each bit corresponds to a fault, "0" means the fault has not occurred, and "1" means the fault has occurred. cc =3 corresponds to "0011" indicating f tube,cc and f fin,cc occurs, and f vlvso,cc and f vlvsc,cc It does not happen, so the design defines the target device status quantity to indicate that the target device is in a normal working state or the target device is in various fault states.

[0104] Step C. Based on the target fault feature data matrix corresponding to the target device, combined with the target fault features corresponding to the target device control quantity and the target device state quantity of each preset historical time node, the EM algorithm is used to train the CIOHMM model corresponding to the target device to obtain the trained CIOHMM model, that is, the fault diagnosis model corresponding to the target device.

[0105] In actual application, the above step C is specifically designed to execute the following steps C1 to C4.

[0106] Step C1. About π={π i}, according to the following formula:

[0107]

[0108] Calculate the initial probability vector π = {π i}, where I 1 Indicates the state of the target device at the first moment, I 1 =i means that the state of the target device corresponding to the first moment is the i-th state, O means the target device corresponds to the target fault feature observation value of all moments, λ = {π,P,b i (O 1 )},b i (O 1 ) represents the probability of each target fault feature observation value of the target device at the first moment corresponding to the i-th state; U represents the control quantity of the target device corresponding to all moments, and Pr(·) represents the probability function.

[0109] Step C2. Regarding the state transfer matrix P = {a ij (U)}, first obtain the following formula:

[0110]

[0111] 1≤k≤K, K represents the total number of moments, I k Indicates the state of the target device at the kth moment, U k represents the control amount of the target device at the kth moment, a 0,ij 、a 1,ij 、a 2,ij represents the constant term, U k Linear coefficient, U k The coefficient of the quadratic term;

[0112] Then derive the following formula:

[0113]

[0114] in, represents the probability of the target device transferring from the state at the kth moment to the state at the k+1th moment under the control quantity at the kth moment, and I represents the state of the target device corresponding to all moments;

[0115] Combination Right now Based on the Lagrange multiplier method, the following function is obtained:

[0116]

[0117] in, represents the Lagrange multiplier, and combined with Then using the chain method, we get:

[0118]

[0119] in,

[0120] First, a ij (U k ) Take the partial derivative and get the following:

[0121]

[0122] Then we further obtain the following:

[0123]

[0124] Summing j on both sides of the equal sign, we get the following:

[0125]

[0126] Substituting the above formula into Obtained as follows:

[0127]

[0128]

[0129] Then we further deduce a 1,ij The iteration formula is as follows:

[0130]

[0131] The same goes for a 2,ij Taking the derivative, we can also get

[0132]

[0133] By using the least squares method, we can estimate a from the above formula 0,ij ,a 1,ij and a 2,ij , i, j = 1, ..., N to obtain the state transfer matrix P. When obtaining the state transfer matrix parameters, there are actually K-1 equations, 3 × N 2 The least squares method can be used to estimate these 3×N unknowns. 2 parameters.

[0134] Step C3. The target fault feature observation values ​​O corresponding to the i-th state of the target device obey the mixed Gaussian distribution as follows:

[0135]

[0136] in:

[0137]

[0138] C represents the number of Gaussian models in the mixed Gaussian distribution, c ic represents the weight of the cth Gaussian model of the mixed normal distribution obeyed by each target fault feature observation value corresponding to the i-th state of the target device, f(O|μ ic ,U ic ) is the probability density function of the cth Gaussian model in the mixed Gaussian distribution corresponding to each target fault feature observation value in the ith state, then the distribution parameter estimate of each target fault feature observation value corresponding to the cth Gaussian model of the target device corresponding to the ith state is The calculation formula is as follows:

[0139]

[0140]

[0141] in:

[0142]

[0143] Step C4. Based on the target fault feature data matrix corresponding to the target device, the target device control quantity and the target device state quantity corresponding to each preset historical time node are respectively matched with the target fault feature. Based on the above iterative formula, each parameter is iteratively updated in turn until the preset convergence condition is met, and the initial probability vector π={π i}、State transfer matrix P={a ij (U)}, observation likelihood function b i (O) Obtain the trained CIOHMM model, that is, the fault diagnosis model corresponding to the target device.

[0144] Based on the fault diagnosis model corresponding to the target device obtained by the above design, in actual application, according to the following step i, the fault diagnosis model corresponding to the target device is applied to realize fault diagnosis of the target device.

[0145] Step i. Obtain the observed value of each target fault feature of the target device at the target time point, combine it with the control value of the target device at the target time point, apply the fault diagnosis model corresponding to the target device, obtain the state value of the target device at the target time point through the Viterbi algorithm, and realize the fault diagnosis of the target device at the target time point.

[0146] In the specific application, it is implemented according to the following steps.

[0147] 1) Initialize the local state, that is, at the initial time t = 1, the hidden state is the maximum probability of all possible state transition paths of i∈{1,2,…,N};

[0148] 2) Perform dynamic programming to recursively calculate the local state at time t=2, 3, ..., K;

[0149] 3) Calculate the probability of the most likely hidden state sequence at time K, and determine the most likely hidden state at time K;

[0150] 4) Use the local state to start backtracking and finally get the most likely hidden state sequence, that is, the estimated state of the cooling coil at different times.

[0151] The above technical scheme designs a fault diagnosis method for a closed-loop control system based on CIOHMM, which uses feedback neural networks, Kalman filtering / particle filtering and other methods to extract fault features, thereby obtaining feature data sets of major fault types and establishing a CIOHMM model for realizing target equipment state quantity judgment; unlike traditional HMM and IOHMM, which have inputs, the continuous closed-loop control signal of the equipment is set as the input of the CIOHMM model, and different state transfer matrices are used to describe the dynamic changes of the state for different inputs, which can describe the state evolution of the equipment under closed-loop control; the CIOHMM model established in this way not only describes the relationship between the equipment state and the observed value, but also describes the impact of closed-loop control on the dynamic changes of the system state, which can realize accurate diagnosis of air-conditioning unit equipment faults, and has unique advantages over other methods.

[0152] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A fault diagnosis method for a closed-loop control system based on CIOHMM, characterized in that: For each target device of the air-conditioning unit, perform the following steps A to C to obtain a fault diagnosis model corresponding to the target device; then according to step i, apply the fault diagnosis model corresponding to the target device to implement fault diagnosis of the target device; Step A. Obtain each target fault feature corresponding to the target device, and obtain the data value of each target fault feature corresponding to each preset historical time node, construct a target fault feature data matrix corresponding to the target device, and then enter step B; the above step A includes the following steps A1 to A4; Step A1. Define the difference between the data of each variable with a set value involved in the target device and its set value, and preset each sensor detection related to the fault involved in the target device as each fault feature to be analyzed, and proceed to step A2; Step A2. Based on the feedback neural network model corresponding to the pre-trained target device, which takes the preset target detection feature data as input and the preset target analysis feature estimation value as output, define the difference between the target analysis feature estimation value output by the feedback neural network model and the target analysis feature calculation value as the fault feature to be analyzed, and then enter step A3; Step A3. Use Kalman filtering or particle filtering to estimate various parameters related to the fault of the target device as the fault characteristics to be analyzed, and then proceed to step A4; Step A4. Based on the data values ​​of each preset historical time node corresponding to each fault feature to be analyzed, the PCA principal component analysis method is used to obtain each principal component fault feature to be analyzed in each fault feature to be analyzed as each target fault feature corresponding to the target device, and the data values ​​of each target fault feature corresponding to each preset historical time node are obtained to construct a target fault feature data matrix corresponding to the target device; Step B. Construct the initial probability vector π corresponding to the target device = {π i }、State transfer matrix P={a ij (U)}, observation likelihood function b i (O), considering the target device control quantity and the target device corresponding to each target fault characteristic observation value, a CIOHMM model for judging the target device state quantity is implemented, wherein the target device control quantity is used to represent the working mode of the target device, and the target device state quantity is used to represent that the target device is in a normal working state or in various fault states, i = 1, ..., N, j = 1, ..., N, N represents the number of states of the target device, π i represents the initial probability of the target device being in the i-th state, a ij (U) represents the probability of transition from the i-th state to the j-th state under the target device control quantity U, O represents the target device corresponding to each target fault characteristic observation value, b i (O) represents the probability of each target fault feature observation value of the target device corresponding to the i-th state; then proceed to step C; Step C. Based on the target fault feature data matrix corresponding to the target device, combined with the target fault features, the target device control quantity and the target device state quantity corresponding to each preset historical time node are respectively corresponded, and the EM algorithm is used to train the CIOHMM model corresponding to the target device to obtain the trained CIOHMM model, that is, the fault diagnosis model corresponding to the target device; Step i. Obtain the observed value of each target fault feature of the target device at the target time point, combine it with the control value of the target device at the target time point, apply the fault diagnosis model corresponding to the target device, obtain the state value of the target device at the target time point through the Viterbi algorithm, and realize the fault diagnosis of the target device at the target time point.

2. A fault diagnosis method for a closed-loop control system based on CIOHMM according to claim 1, characterized in that: If the target device is a cooling coil, in step A2, based on the pre-trained cooling coil corresponding to the chilled water flow Mixed air flow Mixed air temperature T a,mix , Chilled water supply temperature T chw,sup is the input and heat transfer Q cc The estimated value is the feedback neural network model with the output, combined with Heat exchange Q cc Calculated value, defining the heat transfer Q output by the feedback neural network model cc Estimated value and heat transfer Q cc The difference between the calculated values ​​is used as the fault feature to be analyzed, where c pw represents the specific heat capacity of chilled water, T chw,rn Indicates the chilled water return temperature.

3. The fault diagnosis method of a closed-loop control system based on CIOHMM according to claim 1 is characterized in that: The step C includes the following steps C1 to C4; Step C1. About π={π i }, according to the following formula: Calculate the initial probability vector π = {π i }, where I1 represents the state of the target device at the first moment, I1=i represents the state of the target device at the first moment is the i-th state, O represents the observed value of each target fault feature of the target device at all moments, λ={π,P,b i (O1)}, b i (O1) represents the probability of each target fault characteristic observation value of the target device at the first moment corresponding to the i-th state; U represents the control amount of the target device corresponding to all moments, and Pr(·) represents the probability function; Step C2. Regarding the state transfer matrix P = {a ij (U)}, first obtain the following formula: 1≤k≤K, K represents the total number of moments, I k Indicates the state of the target device at the kth moment, U k represents the control amount of the target device at the kth moment, a 0,ij 、a 1,ij 、a 2,ij represents the constant term, U k Linear coefficient, U k The coefficient of the quadratic term; Then derive the following formula: in, represents the probability of the target device transferring from the state at the kth moment to the state at the k+1th moment under the control quantity at the kth moment, and I represents the state of the target device corresponding to all moments; Combination Right now Based on the Lagrange multiplier method, the following function is obtained: in, represents the Lagrange multiplier, and combined with Then using the chain method, we get: in, First, a ij (U k ) Take the partial derivative and get the following: Then we further obtain the following: Summing j on both sides of the equal sign, we get the following: Substituting the above formula into Obtained as follows: Then we further deduce a 1,ij The iteration formula is as follows: Similarly for a 2,ij Taking the derivative, we can also get By using the least squares method, we can estimate a from the above formula 0,ij ,a 1,ij and a 2,ij ,i,j=1,…,N to obtain the state transfer matrix P; Step C3. The target fault feature observation values ​​O corresponding to the i-th state of the target device obey the mixed Gaussian distribution as follows: in: C represents the number of Gaussian models in the mixed Gaussian distribution, c ic represents the weight of the cth Gaussian model of the mixed normal distribution obeyed by each target fault feature observation value corresponding to the i-th state of the target device, f(O|μ ic ,U ic ) is the probability density function of the cth Gaussian model in the mixed Gaussian distribution corresponding to each target fault feature observation value in the ith state, then the distribution parameter estimate of each target fault feature observation value corresponding to the cth Gaussian model of the target device corresponding to the ith state is The calculation formula is as follows: in: a k (i)=Pr(O1O2…O k ,I k =iλ) β k (i)=Pr(O k+1 O k+2 …O K |I k =i,λ) Step C4. Based on the target fault feature data matrix corresponding to the target device, the target device control quantity and the target device state quantity corresponding to each preset historical time node are respectively matched with the target fault feature. Based on the above iterative formula, each parameter is iteratively updated in turn until the preset convergence condition is met, and the initial probability vector π={π i }、State transfer matrix P={a ij (U)}, observation likelihood function b i (O) Obtain the trained CIOHMM model, that is, the fault diagnosis model corresponding to the target device.

Images