A method for air conditioning fault diagnosis based on residual multidimensional Taylor net

Abstract

The application relates to a heating ventilation air conditioner fault diagnosis method based on a residual multidimensional Taylor network, aiming at each type of fault corresponding to a target heating ventilation air conditioner, considering each heating ventilation air conditioner variable type data, training a residual multidimensional Taylor network (ResMTN) classifier newly designed based on a BP-MTN classifier, and obtaining a heating ventilation air conditioner fault diagnosis model, wherein the ResMTN classifier sequentially comprises an input layer, a polynomial layer, an activation layer, a residual connection module, a full connection layer, a Softmax layer and an output layer, and then the fault diagnosis of the target heating ventilation air conditioner is realized in application; the design scheme simplifies the model complexity, improves the model generalization ability, compared with common machine learning classifiers, can improve the detection efficiency of the heating ventilation air conditioner while not reducing the classification accuracy.

Description

TECHNICAL FIELD

[0001] The application relates to a residual multi-dimensional Taylor net-based air conditioning fault diagnosis method and belongs to the technical field of air conditioning fault detection. BACKGROUND

[0002] Building energy consumption, industrial energy consumption and transportation energy consumption are the three major energy consumers in China. According to statistics, building energy consumption accounts for 28.3% of total social electricity consumption, and air conditioning electricity consumption accounts for more than 60% of building electricity consumption.

[0003] About 42% of refrigeration energy consumption and 26% of maintenance costs in air conditioning are caused by equipment failure. It is estimated that air conditioning failure will cause energy waste and failure to meet the comfort needs of people. Due to high diagnosis accuracy, deep learning methods are widely used in the field of air conditioning fault diagnosis. However, these deep models have complex structures, long model training time and large calculation amount, and may not meet the needs of rapid fault diagnosis. In addition, due to the complexity of existing models, there may be problems such as overfitting and gradient disappearance. SUMMARY

[0004] The technical problem to be solved by the application is to provide a residual multi-dimensional Taylor net-based air conditioning fault diagnosis method, which improves the detection efficiency of air conditioning through innovative classifier structure design.

[0005] The application adopts the following technical solutions to solve the above technical problems: the application designs a residual multi-dimensional Taylor net-based air conditioning fault diagnosis method, executes steps A to D to obtain an air conditioning fault diagnosis model, and then executes step i to apply the air conditioning fault diagnosis model to complete fault diagnosis of the target air conditioning at the target detection time.

[0006] Step A. Based on the target air conditioning corresponding to each historical sample time of each type of preset fault, data of each type of preset air conditioning variable of the target air conditioning corresponding to each historical sample time are obtained, and then step B is entered.

[0007] Step B. The type fault corresponding to the historical sample time of the target air conditioning and the data of each type of air conditioning variable constitute a sample, and each sample is obtained, and then step C is entered.

[0008] Step C. Based on the BP-MTN classifier, a ResMTN classifier is constructed for inputting the data of each type of air conditioning variable of the target air conditioning at the corresponding time and outputting the type fault of the target air conditioning at the corresponding time, and then step D is entered.

[0009] Step D. Based on each sample, take the data of each HVAC variable type corresponding to the historical sample time of the target HVAC in the sample as input, and take the type of fault of the target HVAC in the sample corresponding to the historical sample time as output, train the ResMTN classifier to obtain the HVAC fault diagnosis model.

[0010] Step i. Collect data on the variable types of each HVAC system corresponding to the target detection time, and apply the HVAC fault diagnosis model to obtain the type of fault corresponding to the target HVAC system at the target detection time.

[0011] As a preferred technical solution of the present invention: In step A, based on the data of each preset HVAC variable type corresponding to each historical sample time of the target HVAC, the data of each historical sample time corresponding to each HVAC variable type is normalized for each HVAC variable type, the data of each historical sample time corresponding to the HVAC variable type is updated, and then the data of each preset HVAC variable type corresponding to each historical sample time of the target HVAC is updated, and then the process proceeds to step B.

[0012] As a preferred technical solution of the present invention: in step A, for each HVAC variable type, further for the data x of each historical sample time corresponding to each HVAC variable type. i (t'), according to the following formula:

[0013]

[0014] Perform normalization to obtain normalized data. Where 1≤i≤I, I represents the number of HVAC variable types, x i (t') represents the data at time t' corresponding to the historical sample of the i-th HVAC variable type. x represents i (t') is the normalized data, min(X) i ) represents the minimum value among the historical sample data corresponding to the i-th HVAC variable type at each time point, max(X i ) represents the maximum value of the data at each historical sample time corresponding to the i-th HVAC variable type, and s represents the preset scaling factor, s∈[0,1].

[0015] As a preferred technical solution of the present invention: the ResMTN classifier constructed in step C consists of an input layer, a polynomial layer, an activation layer, a residual connection module, a fully connected layer, a Softmax layer, and an output layer connected in series from its input end to its output end, and the output end of the input layer is simultaneously connected to the input end of the residual connection module.

[0016] As a preferred technical solution of the present application: the input layer in the ResMTN classifier constructed in step C is used to receive data X(t) = x1(t), …, x i (t), …, x I (t) of each HVAC variable type of the target HVAC at time t, and forward the polynomial layer, 1≤i≤I, I represents the number of HVAC variable types, x i (t) represents the data of the i-th HVAC variable type of the target HVAC at time t.

[0017] The polynomial layer constructs the output Y(k) of the polynomial layer according to the preset highest expansion term M corresponding to the network to be trained, according to the data X(t) of each HVAC variable type of the target HVAC at time t, and transmits it to the activation layer, 1≤M≤I;

[0018] Y(t) = [y1(t) … y i (t) … y I (t)] = [W1 … W i … W I ] T ·V(t)

[0019] Wherein, y i (t) represents the output of the polynomial layer about the i-th HVAC variable type of the target HVAC at time t, W i = [w i，1 … w i，l … w i，L(I，M) ], 1≤l≤L(I, M), L(I, M) represents the number of monomials in the polynomial based on I and M, W i represents the weight vector of the i-th HVAC variable type of the target HVAC in the polynomial layer, w i，l represents the l-th weight in W i , w i，l is a parameter to be trained in the ResMTN classifier, V(t) = [1, x1(t) … x I (t), x1(t)x2(t) … (x I (t)) 2 , …, x1(t) … x M (t) … (x I (t)) M ] T , V(t) represents the expansion of each term of the data X(t) of each HVAC variable type of the target HVAC at time t about the highest expansion term M.

[0020] As a preferred technical solution of the present application: the activation layer in the ResMTN classifier constructed in step C is the activation function Tanh, the activation function Tanh is used to receive the output Y(t)=[y1(t) … y i (t) … y I (t)] from the polynomial layer, the output A(t) of the activation layer is constructed as follows, and is transmitted to the residual connection module;

[0021] A(t)=tanh[Y(t)|=|a1(t) … a i (x) … a I (t)|

[0022] Wherein, I represents the number of HVAC variable types, a i (t) represents the output of the i-th HVAC variable type of the target HVAC at time t by the activation function Tanh, y i (t) represents the output of the i-th HVAC variable type of the target HVAC at time t by the polynomial layer, exp(y i (t)) represents e to the power of y i (t), exp(-y i (t)) represents e to the power of-y i (t).

[0023] As a preferred technical solution of the present application: the residual connection module in the ResMTN classifier constructed in step C is used to receive the output A(t) from the activation layer and the output X(t) from the input layer respectively, the output R(t) of the residual connection module is constructed as R(t)=A(t)+X(t), and is transmitted to the fully connected layer.

[0024] As a preferred technical solution of the present application: the fully connected layer in the ResMTN classifier constructed in step C constructs the output Z(t) of the fully connected layer as follows for R(t) output from the residual connection module, and transmits to the Softmax layer;

[0025] Z(t)=Λ T ·R(t)=[Λ1 … Λ c … Λ C ] T ·R(t)=[z1(t) … z c (t) … z C (t)]

[0026] Wherein, 1≤c≤C, C represents the type number of the preset types of faults, Λ c represents the weight column vector of the c-th type of fault, λc，i represents the weight value of the i-th HVAC variable type to the c-th type fault, λ c，i is a parameter to be trained in the ResMTN classifier, z c (t) represents the output of the fully connected layer corresponding to the c-th type fault of the target HVAC at time t.

[0027] As a preferred technical solution of the present application: the Softmax layer in the ResMTN classifier constructed in step C outputs the probabilities P(t) of the target HVAC corresponding to each type of fault at time t from Z(t) = [z1(t) … z c (t) … z C (t)] output by the fully connected layer, and transmits to the output layer, and the output layer outputs the type of fault corresponding to the target HVAC at time t according to P(t).

[0028] P(t) = [p1(t) … p c (t) … p C (t)] T

[0029] wherein, C represents the type number of each type of fault, z c (t) represents the output of the fully connected layer corresponding to the c-th type fault of the target HVAC at time t, exp(z c (t)) represents e raised to the power of z c (t), and p c (t) represents the probability of the Softmax layer outputting the c-th type fault of the target HVAC at time t.

[0030] As a preferred technical solution of the present application: in the training process of the ResMTN classifier, the cross-entropy function is as follows:

[0031]

[0032] The Adam algorithm is applied to train the ResMTN classifier to obtain the HVAC fault diagnosis model, wherein L represents the loss result, 1≤c≤C, c represents the type number of each type of fault, p c (t) represents the probability of the Softmax layer outputting the c-th type fault of the target HVAC at time t in the ResMTN classifier, h c (t) represents the true label of whether the target HVAC corresponds to the c-th type fault at time t.

[0033] The HVAC fault diagnosis method based on the residual multidimensional Taylor network has the following technical effects compared with the prior art.

[0034] (1) The HVAC fault diagnosis method based on the residual multidimensional Taylor network is designed, each type of fault corresponding to the target HVAC is considered, each HVAC variable type data is considered, the ResMTN classifier based on the BP-MTN classifier is trained, and an HVAC fault diagnosis model is obtained, wherein the ResMTN classifier sequentially includes an input layer, a polynomial layer, an activation layer, a residual connection module, a full connection layer, a Softmax layer, and an output layer, and then the fault diagnosis of the target HVAC is realized in application; the design scheme simplifies the model complexity, improves the model generalization ability, and compared with the commonly used machine learning classifier, the detection efficiency of the HVAC can be improved without reducing the classification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a structural schematic diagram of the ResMTN classifier in the design of the present application;

[0036] Figure 2 is a confusion matrix schematic diagram obtained by the fault diagnosis in the actual application of the design of the present application;

[0037] Figure 3 is a convergence schematic diagram of two-order ResMTN, CNN, TCN, and LSTM classifiers in the implementation application. DETAILED DESCRIPTION

[0038] The specific embodiments of the present application will be further described in detail below in combination with the drawings of the specification.

[0039] The HVAC fault diagnosis method based on the residual multidimensional Taylor network is designed, and in the actual application, the following steps A to step D are specifically designed and executed to obtain an HVAC fault diagnosis model.

[0040] Step A. Based on each historical sample time corresponding to each type of fault of the target HVAC, the preset each HVAC variable type corresponding to each historical sample time of the target HVAC is obtained, including mixed air temperature T a,mix , supply air humidity W a,sup , return air humidity W a,rn , fresh air humidity W a,oa , chilled water flow M chw , return air flow M a,rn , chilled water supply water temperature T chw,sup , fresh air flow M a,oa , cooling coil outlet air temperature T a,dis,cc , and supply air fan power Qsf data of each historical sample moment of the HVAC variable type corresponding to the HVAC variable type, and then update the data of each historical sample moment of the preset each HVAC variable type of the target HVAC respectively corresponding to each historical sample moment of the HVAC, and then enter step B.

[0041] Here, the normalization operation involved in step A is, in actual application, further respectively for each HVAC variable type, and further respectively for the data x i (t') of each historical sample moment of the HVAC variable type corresponding to the HVAC variable type, according to the following formula:

[0042]

[0043] Perform normalization processing to obtain normalized data Wherein, 1≤i≤I, I represents the number of HVAC variable types, x i (t') represents the data of the i-th HVAC variable type corresponding to the historical sample moment t' of the HVAC, x i (t') normalized data, min(X i ) represents the minimum value of the data of the i-th HVAC variable type corresponding to each historical sample moment, max(X i ) represents the maximum value of the data of the i-th HVAC variable type corresponding to each historical sample moment, and s represents a preset scaling coefficient, s∈[0,1], that is, the above normalization formula scales the input data to the range of [s,1], which has a smaller difference between the mean value and the variance of the high-order expansion term and the low-order expansion term, while maintaining a sufficient range for the decision boundary to learn.

[0044] Step B. Form samples with the type fault of the target HVAC corresponding to the historical sample moment and the data of each HVAC variable type, obtain each sample, and then enter step C.

[0045] Step C. Based on the BP-MTN classifier, construct a ResMTN classifier for applying the data of each HVAC variable type of the target HVAC corresponding to the moment as input and the type fault of the target HVAC corresponding to the moment as output, and then enter step D.

[0046] The ResMTN classifier designed here is sequentially connected in the direction from its input end to its output end as input layer, polynomial layer, activation layer, residual connection module, full connection layer, Softmax layer, and output layer, and the output end of the input layer is also connected to the input end of the residual connection module.

[0047] Wherein, the input layer is used for receiving data X(t) = x1(t), …, xI(t) of each HVAC variable type of the target HVAC at time t, and forwarding the polynomial layer, 1≤i≤I, I represents the number of HVAC variable types, x i (t) represents the data of the i-th HVAC variable type of the target HVAC at time t. I (t), and forwarding the polynomial layer, 1≤i≤I, I represents the number of HVAC variable types, x i (t) represents the data of the i-th HVAC variable type of the target HVAC at time t.

[0048] The polynomial layer constructs the output Y(k) of the polynomial layer according to the preset highest expansion term M corresponding to the to-be-trained network based on the received data X(t) of each HVAC variable type of the target HVAC at time t, and transmits to the activation layer, 1≤M≤I.

[0049] Y(t) = [y1(t) … y i (t) … y I (t)] = [W1 … W i … W I ] T ·V(t)

[0050] Wherein, y i (t) represents the output of the polynomial layer about the i-th HVAC variable type of the target HVAC at time t, W i = [w i，1 … w i，l … w i，L(I，M) ], 1≤l≤L(I, M), L(I, M) represents the number of monomials in the polynomial based on I and M, W i represents the weight vector of the i-th HVAC variable type of the target HVAC in the polynomial layer, w i，l represents the l-th weight in W i , w i，l is a to-be-trained parameter in the ResMTN classifier, V(t) = [1, x1(t), …, x I (t), x1(t)x2(t), …, (x I (t)) 2 , …, x1(t), …, x M (t), …, (x I (t)) M ] T , V(t) represents the expansion of each term of the data X(t) of each HVAC variable type of the target HVAC at time t about the highest expansion term M.

[0051] In actual sample detection, the number of air conditioning variable types is 10, the type number C of each type of fault is 13, and when the preset highest expansion term M of the polynomial layer is 2, the optimal diagnosis result can be obtained through experiment verification.

[0052] In order to improve the accuracy of the model, the application introduces an activation function at the output of the polynomial layer. After experiment, it is found that the activation function Tanh performs best, therefore, the activation layer in the constructed ResMTN classifier is the activation function Tanh, the activation function Tanh is used to receive the output Y(t)=[y1(t) … y i (t) … y I (t)] from the polynomial layer, and the output A(t) of the activation layer is constructed as follows and transmitted to the residual connection module;

[0053] A(t)=tanh[Y(t)]=[a1(t) … a i (x) … a I (t)]

[0054] Wherein, I represents the number of air conditioning variable types, a i (t) represents the output of the activation function Tanh about the i-th air conditioning variable type of the target air conditioning at time t, y i (t) represents the output of the polynomial layer about the i-th air conditioning variable type of the target air conditioning at time t, exp(y i (t)) represents e to the power of y i (t), exp(-y i (t)) represents e to the power of-y i (t).

[0055] The residual connection module is used to receive the output A(t) from the activation layer and the output X(t) from the input layer, construct the output R(t) of the residual connection module according to R(t)=A(t)+X(t), and transmit it to the full connection layer.

[0056] Considering the output dimension of the designed ResMTN classifier, such as the preset type number C of each type of fault is 13, the full connection layer after the residual connection module connects 10-dimensional R(k) to 13-dimensional, to ensure that the input and output of the softmax layer are the same dimension, therefore, the full connection layer constructs the output Z(t) of the full connection layer according to R(t) from the output of the residual connection module as follows, and transmits it to the Softmax layer;

[0057] Z(t)=Λ T ·R(t)=[Λ1 … Λ c … ΛC ] T • R(t) = [z1(t) … z c (t) … z C (t)]

[0058] wherein, C represents the type number of preset each type of fault, Λ c represents the weight column vector of the cth type of fault, λ c,i represents the weight of the ith HVAC variable type to the cth type of fault, λ c,i is the to-be-trained parameter in the ResMTN classifier, z c (t) represents the output of the fully connected layer about the target HVAC corresponding to the cth type of fault at time t.

[0059] The Softmax layer outputs the probability P(t) of the target HVAC corresponding to each type of fault at time t according to Z(t) = [z1(t) … z c (t) … z C (t)] from the output of the fully connected layer, and transmits to the output layer, and the output layer outputs the type of fault corresponding to the target HVAC at time t according to P(t);

[0060] P(t) = [p1(t) … p c (t) … p C (t)] T

[0061] wherein, C represents the type number of preset each type of fault, z c (t) represents the output of the fully connected layer about the target HVAC corresponding to the cth type of fault at time t, exp(z c (t)) represents e raised to the power of z c (t), and p c (t) represents the probability of the Softmax layer outputting the target HVAC corresponding to the cth type of fault at time t.

[0062] Step D. Based on each sample, taking the data of each HVAC variable type corresponding to the target HVAC at the time of the historical sample in the sample as input, and taking the type of fault corresponding to the target HVAC at the time of the historical sample in the sample as output, according to the following cross-entropy function:

[0063]

[0064] The Adam algorithm is applied to train the ResMTN classifier to obtain the HVAC fault diagnosis model, wherein L represents the loss result, 1≤c≤C, C represents the type number of preset each type of fault, and pc (t) represents the probability of the target HVAC corresponding to the cth type of fault at time t output by the Softmax layer in the ResMTN classifier, h c (t) represents the true label of whether the target HVAC corresponds to the cth type of fault at time t.

[0065] Based on the HVAC fault diagnosis model obtained above, the following steps i are further performed in actual application to apply the HVAC fault diagnosis model to complete fault diagnosis of the target HVAC corresponding to the target detection time.

[0066] Step i. Collect data of each HVAC variable type of the target HVAC corresponding to the target detection time, and apply the HVAC fault diagnosis model to obtain the type of fault corresponding to the target HVAC at the target detection time.

[0067] The above design is applied in actual application to diagnose the 13 types of faults for each HVAC variable type of the target HVAC corresponding to the mixed air temperature T a,mix , the supply air humidity W a,sup , the return air humidity W a,rn , the fresh air humidity W a,oa , the chilled water flow M chw , the return air flow M a,rn , the chilled water supply temperature T chw,sup , the fresh air flow M a,oa , the outlet air temperature T a,dis,cc of the cooling coil, and the power Q sf of the supply air fan based on the HVAC fault diagnosis model obtained by training the ResMTN classifier according to the above design.

[0068] In actual application, the fault diagnosis capability of the ResMTN classifier is verified by using the actual data of ASHRAE-1312. The data of ASHRAE-1312 contains 13 types of faults. Figure 2 The confusion matrix of the average fault diagnosis result of the two-order ResMTN classifier obtained by ten-fold cross-validation is shown in Table 1. Since the test set has 702 samples, a total of 13 types of faults, and 54 samples for each class. It can be seen that 53.9 out of 54 samples of F12 are correctly identified. The decimal number appears here because the average value is taken. The diagnosis accuracy of other faults can reach 100%. Figure 3The convergence of the two-order ResMTN, CNN, TCN and LSTM classifiers is shown, and it can be seen from the figure that the ResMTN classifier converges quickly, but has the highest convergence accuracy, because the design scheme simplifies the model complexity and improves the model generalization ability, so compared with the commonly used machine learning classifiers, the detection efficiency of the heating ventilation air conditioner can be improved without reducing the classification accuracy.

[0069] The embodiments of the present application are described in detail above with reference to the drawings, but the present application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the present application.

Claims

1. A method for HVAC fault diagnosis based on residual multidimensional Taylor networks, characterized in that: Perform steps A to D to obtain the HVAC fault diagnosis model, and then perform step i to apply the HVAC fault diagnosis model to complete the fault diagnosis of the target HVAC at the corresponding target detection time. Step A. Based on the historical sample time of each preset type of fault corresponding to the target HVAC, obtain the data of each preset HVAC variable type corresponding to each historical sample time of the target HVAC, and then proceed to Step B; Step B. Use the type of fault corresponding to the target HVAC system at the historical sample time and the data of each HVAC variable type to form a sample, obtain each sample, and then proceed to step C; Step C. Based on the BP-MTN classifier, construct a ResMTN classifier that takes the data of each HVAC variable type at the corresponding time of the target HVAC as input and the type of fault of the target HVAC at the corresponding time as output, and then proceed to step D; Step D. Based on each sample, take the data of each HVAC variable type corresponding to the historical sample time of the target HVAC in the sample as input, and take the type of fault of the target HVAC in the sample corresponding to the historical sample time as output, train the ResMTN classifier to obtain the HVAC fault diagnosis model. The ResMTN classifier constructed in step C above consists of an input layer, a multinomial layer, an activation layer, a residual connection module, a fully connected layer, a softmax layer, and an output layer, connected in series from its input end to its output end. The output end of the input layer is also connected to the input end of the residual connection module. The input layer of the ResMTN classifier is used to receive data X(t) = x1(t), ..., x2(t) of various HVAC variable types at time t corresponding to the target HVAC system. i (t), ..., x I (t), and forward to the polynomial layer, 1≤i≤I, where I represents the number of HVAC variable types, x i (t) represents the data of the i-th HVAC variable type corresponding to time t of the target HVAC system; For the data X(t) of each HVAC variable type received at time t of the target HVAC, the polynomial layer constructs the output Y(k) of the polynomial layer according to the preset highest expansion term M corresponding to the network to be trained, as follows, and transmits it to the activation layer, 1≤M≤I; Y(t)=[y1(t) … y i (t) … y I (t)]=[W1 … W i … W I (t)] T ·V(t) Among them, y i (t) represents the output of the polynomial layer with respect to the i-th HVAC variable type at time t corresponding to the target HVAC, W i =[w i,1 … w i,l … w i,L(I,M) ], 1≤l≤L(I,M), where L(I,M) represents the number of monomials in the polynomial based on I and M, W i w represents the weight vector of the i-th HVAC variable type in the polynomial layer concerning the target HVAC system. i,l W i The l-th weight, w i,l These are the training parameters in the ResMTN classifier, V(t) = [1, x1(t)...x1(t)]. I (t), x1(t)x2(t)…(x I (t)) 2 ,…,x1(t)…x M (t)…(x I (t)) M ] T V(t) represents the expansion of each HVAC variable type data X(t) with respect to the highest expansion term M at time t of the target HVAC system; Step i. Collect data on the variable types of each HVAC system corresponding to the target detection time, and apply the HVAC fault diagnosis model to obtain the type of fault corresponding to the target HVAC system at the target detection time.

2. The HVAC fault diagnosis method based on residual multidimensional Taylor networks according to claim 1, characterized in that: In step A, based on the data of each preset HVAC variable type corresponding to each historical sample time of the target HVAC, the data of each historical sample time corresponding to each HVAC variable type is normalized for each HVAC variable type, the data of each historical sample time corresponding to the HVAC variable type is updated, and then the data of each preset HVAC variable type corresponding to each historical sample time of the target HVAC is updated, and then the process proceeds to step B.

3. The HVAC fault diagnosis method based on residual multidimensional Taylor networks according to claim 2, characterized in that: In step A, for each HVAC variable type, further analysis is performed on the data x corresponding to each historical sample time for each HVAC variable type. i (t′), according to the following formula: Perform normalization to obtain normalized data. Where 1≤i≤I, I represents the number of HVAC variable types, x i (t′) represents the data at time t′ corresponding to the historical sample of the i-th HVAC variable type. x represents i (t′) is the normalized data, min(X) i ) represents the minimum value among the historical sample data corresponding to the i-th HVAC variable type at each time point, max(X i ) represents the maximum value of the data at each historical sample time corresponding to the i-th HVAC variable type, and s represents the preset scaling factor, s∈[0,1].

4. The HVAC fault diagnosis method based on residual multidimensional Taylor networks according to claim 1, characterized in that: In step C, the activation layer of the ResMTN classifier is the activation function Tanh. The activation function Tanh is used to receive the output Y(t) = [y1(t) … y2(t)] from the multinomial layer. i (t) … y I The output A(t) of the activation layer is constructed as follows and transmitted to the residual connection module; A(t)=tanh[Y(t)]=[a1(t) … a i (t) … a I (t)] in, I represents the number of HVAC variable types, a i (t) represents the output of the activation function Tanh with respect to the i-th HVAC variable type at time t corresponding to the target HVAC system, y i (t) represents the output of the polynomial layer with respect to the i-th HVAC variable type at time t corresponding to the target HVAC, exp(y i (t) represents the y-value of e. i (t) raised to the power of exp(-y) i (t) represents the -y of e. i (t)th power.

5. The HVAC fault diagnosis method based on residual multidimensional Taylor networks according to claim 1, characterized in that: In step C, the residual connection module in the ResMTN classifier is used to receive the output A(t) from the activation layer and the output X(t) from the input layer, respectively. The output R(t) of the residual connection module is constructed according to R(t) = A(t) + X(t) and transmitted to the fully connected layer.

6. The HVAC fault diagnosis method based on residual multidimensional Taylor networks according to claim 1, characterized in that: In step C, the fully connected layer of the ResMTN classifier is constructed based on the output R(t) from the residual connection module. The output Z(t) of the fully connected layer is constructed as follows and transmitted to the Softmax layer. Z(t)=Λ T ·R(t)=[Λ1 … Λ c …Λ C ] T ·R(t)=pz1(t) … z c (t) … z C (t)] Among them, Λ c =[λ c,1 … λ c,i … λ c,I ] T 1≤c≤C, where C represents the number of preset fault types, Λ c λ represents the weight column vector of the c-th type of fault. c,i λ represents the weight from the i-th HVAC variable type to the c-th type of fault. c,i These are the training parameters in the ResMTN classifier, z c (t) represents the output of the fully connected layer regarding the c-th type of fault in the target HVAC system at time t.

7. The HVAC fault diagnosis method based on residual multidimensional Taylor networks according to claim 1, characterized in that: In step C, the Softmax layer of the ResMTN classifier is designed for the output Z(t) = [z1(t)... z...] from the fully connected layer. c (t)… z C The probability P(t) of each type of fault of the target HVAC at time t is output by the Softmax layer as follows, and transmitted to the output layer. The output layer outputs the type of fault of the target HVAC at time t based on P(t). P(t)=[p1(t) … p c (t) … p C (t)] T in, 1≤c≤C, where C represents the preset number of fault types, z c (t) represents the output of the fully connected layer regarding the target HVAC system at time t, corresponding to the Cth type of fault, exp(z c (t) represents the z-value of e. c (t)th power, p c (t) represents the probability of the target HVAC system output by the Softmax layer corresponding to the c-th type of fault at time t.

8. The HVAC fault diagnosis method based on residual multidimensional Taylor networks according to claim 1, characterized in that: In step D, during the training of the ResMTN classifier, the following cross-entropy function is used: The Adam algorithm is applied to train the ResMTN classifier to obtain an HVAC fault diagnosis model, where L represents the loss result, 1≤c≤C, C represents the preset number of each fault type, and p c (t) represents the probability of the target HVAC system corresponding to the c-th type of fault at time t, output by the Softmax layer in the ResMTN classifier, h. c (t) indicates whether the target HVAC system corresponds to the true label of the c-th type of fault at time t.

Images