Fault diagnosis method of dynamic multi-dimensional Taylor network based on gradient scoring

Abstract

The invention discloses a fault diagnosis method for a dynamic multi-dimensional Taylor network based on gradient scoring, and the method comprises the steps: firstly constructing a fault signal feature data set which comprises a plurality of fault original signals and corresponding fault labels; and then constructing a fault diagnosis neural network of a dynamic multi-dimensional Taylor network based on gradient scoring, and performing forward propagation on input data through an input layer, a full connection layer, a polynomial layer, a Sigmoid layer, a gradient layering layer, a scoring sorting selection layer and an output layer in sequence. Calculating the gradient of a loss function relative to each monomial coefficient in the gradient layering by using an automatic differential technology, sorting monomials in the scoring sorting selection layer according to the gradient, dynamically adjusting the activation state of the monomials, carrying out backward propagation according to the calculated gradient, updating parameters of each layer of the model, and carrying out grading sorting on the monomials. And carrying out backward propagation according to the calculated gradient, updating parameters of each layer of the model, and finally, outputting a fault diagnosis result by an output layer.

Description

Technical Field

[0001] The invention belongs to the field of machine fault diagnosis, and in particular relates to a fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring. Background Art

[0002] In the fault diagnosis of air handling units (AHU) in air conditioning systems, multidimensional Taylor networks (MTNs) provide an effective tool for approximating nonlinear functions. However, although MTNs have powerful approximation capabilities in theory, they still have some limitations and defects in practical applications, especially in AHU fault diagnosis: (1) Complexity of high-order polynomials: When dealing with highly nonlinear problems, in order to achieve high accuracy, traditional MTNs need to use high-order polynomials. This makes the model very complex and the amount of calculation increases significantly, thereby extending the calculation time. (2) Difficulty in selecting hyperparameters: The highest order of the MTN expansion term is a hyperparameter, and its determination is usually based on experiments and experience. This process is complex and lacks systematicity. It is difficult to guarantee the optimal effect under different data sets and application scenarios. Summary of the invention

[0003] In view of the problems existing in the prior art, the present invention provides a fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring, which can better realize accurate diagnosis of machine faults in different scenarios.

[0004] To solve the above technical problems, the present invention provides the following technical solutions: a fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring, comprising the following steps:

[0005] S1. Based on the original signal of equipment fault, a fault signal feature dataset is constructed, including several original fault signals and corresponding fault labels;

[0006] S2. Construct a fault diagnosis neural network based on a dynamic multidimensional Taylor network with gradient scoring, train the fault diagnosis neural network with the original fault signal as input and the corresponding fault type as output, and obtain a fault diagnosis model; the fault diagnosis neural network includes an input layer, a first polynomial layer, a gradient scoring layer, a scoring sorting selection layer, a second polynomial layer, a first fully connected layer, a softmax layer, a second fully connected layer, a Sigmoid layer, and an output layer;

[0007] Extract fault-related variables from the original fault signal, and then extract fault feature data. The input layer receives the fault feature data, generates and combines the fault feature data into different monomials through the first polynomial layer, and then inputs the monomials into the gradient scoring layer. The gradient of the loss function relative to each monomial coefficient is calculated using automatic differentiation technology. The scoring and sorting selection layer is connected to the gradient scoring layer to sort the monomials according to the calculated gradient scores.

[0008] On the other hand, the input layer is sequentially connected to the first fully connected layer and the Sigmoid layer, and the Sigmoid layer is connected to the scoring and sorting selection layer. The monomials of the polynomial layer are screened according to the proportion of the selected monomials calculated from the gradient scores that represent the nonlinear function.

[0009] The input end of the second polynomial layer is connected to the output end of the scoring and sorting selection layer, the output end of the second polynomial layer is connected to the input end of the second fully connected layer, and the output end of the second fully connected layer is connected to the input end of the output layer of the Softmax layer to obtain the probability of each type. The Softmax layer is connected to the output layer to output the final classification result;

[0010] S3. Collect the original signal of equipment failure, input it into the fault diagnosis model, and obtain the fault classification result.

[0011] Furthermore, in the aforementioned step S2, the first polynomial layer receives an input vector x=(x1, x2, ..., x m ), where m is the dimension of the input, and the output z is generated by combining different monomials, where the monomials include various power combinations of the input variables and cross terms;

[0012] The first polynomial layer is a polynomial representation composed of monomials of different powers, which is used to approximate nonlinear functions. The output z of the first polynomial layer is (z1, z2, ..., z k ), k represents the dimension of the polynomial layer output, where each element z j The calculation formula is:

[0013]

[0014] in, Represents the monomial corresponding to the j-th output dimension The coefficient of n represents the number of input variables in the monomial. l ,l=1,2...,m represents the power of the corresponding input variable.

[0015] Furthermore, the aforementioned gradient grading layer is configured to perform the following steps:

[0016] S2A.1. Define the cross entropy loss function:

[0017]

[0018] in, is the output of the model, and the true label is y=(y1,y2,...,y K ), there are K fault categories.

[0019] S2A.2. Calculate the loss function L relative to the model output Gradient

[0020] According to the derivation formula get:

[0021]

[0022] When i=j,

[0023]

[0024] When i≠j,

[0025]

[0026] Right now

[0027]

[0028] S2A.3. Computational Model Output Relative to the polynomial output z j Gradient Model Output Where f is the Softmax activation function, then for the Softmax activation function,

[0029]

[0030] Furthermore, the above relative monomial coefficient The gradient is as follows: The second polynomial output is transformed to obtain the model output Let the monomial of the polynomial be where x i is the input variable, n k is the power of the corresponding variable, m represents the number of variables in the monomial, and its coefficient is

[0031] And the polynomial output is

[0032]

[0033] According to the polynomial derivation rule,

[0034]

[0035] According to the chain rule, the loss function L is relative to the monomial coefficient The gradient of is:

[0036]

[0037] According to the polynomial derivation rule, we can calculate:

[0038]

[0039] The number of monomials is M, let is the coefficient of the lth monomial, then

[0040]

[0041] Furthermore, the aforementioned input layer is sequentially connected to the first fully connected layer and the Sigmoid layer, and the Sigmoid layer is connected to the scoring sorting selection layer. Let the input be x = (x1, x2, ..., x m ), the weight matrix of the fully connected layer is Let the bias vector be 0, then the output of the first fully connected layer is calculated as:

[0042] s=Ω T x

[0043] The polynomial layer has M monomials. After the scoring and sorting selection layer is sorted according to the gradient score, the selection probability corresponding to the i-th monomial is p i , 0≤p i ≤1, the probability is obtained through the Sigmoid function,

[0044] The output s of the first fully connected layer is the output obtained by the Sigmoid function Each element p i ,i=1,2,...,m is calculated as follows:

[0045]

[0046] The values ​​in these outputs are between 0 and 1 and serve as ratios for selecting the monomials in the polynomial.

[0047] Furthermore, the aforementioned scoring sorting selection layer is connected to the gradient scoring layer to sort the monomials according to the calculated gradient scores, specifically including the following sub-steps:

[0048] S2B.1. Suppose the polynomial layer has M monomials, whose coefficients are w in , i represents the monomial index, n represents the order-related identifier of the monomial, and the loss function calculated at the gradient level is relative to the i-th monomial coefficient w in The gradient is g i ,

[0049] S2B.2, Sorting operation: Define a sorting function S, which sorts the gradient vector g = (g1, g2, ..., g M) to get the sorted index vector k=(k1,k2,…,k M ), so that g k1 ≥g k2 ≥…≥g kM , that is, k = S(g);

[0050] S2B.3. Select monomials based on the sorting results: Let the selected proportion be p i (0 <p i ≤1), then the number of monomials selected is

[0051] S2B.3. Selected monomial index k selected =(k1,k2,…,k N ), that is, the scoring sorting selection layer selects the monomial index according to the gradient score, and the subsequent model performs further calculations based on the selected monomial index.

[0052] Furthermore, the output end of the aforementioned second fully connected layer is connected to the input end of the Softmax layer output layer to obtain the probability of each type, specifically:

[0053] The output of the second fully connected layer is input to the softmax layer. The output of the softmax layer is P = (p1, p2, ..., p K ), where p k represents the probability of belonging to the kth class, and the calculation formula is:

[0054]

[0055] Through the Softmax function, the output of the fully connected layer is converted into a probability distribution of each type.

[0056] Furthermore, the aforementioned Softmax layer is connected to the output layer to output the final classification result, which is:

[0057] The output of the output layer is finally the classification result y, and its formula is:

[0058]

[0059] Where k = 1, 2, ..., K. That is, the output layer selects the category index with the largest vector probability in the probability vector output by the Softmax layer as the final classification result.

[0060] Furthermore, the monomials of the aforementioned first polynomial layer and the second monomial layer include but are not limited to 1st order, 2nd order, 3rd order and higher order monomials, and the coefficient of each monomial is an adjustable parameter.

[0061] Compared with the prior art, the beneficial technical effects of the above technical solutions adopted by the present invention are as follows: The fault diagnosis method of the dynamic multidimensional Taylor network based on gradient scoring provided by the present invention is aimed at the fault diagnosis of the air handling unit in the air conditioning unit, and the specific beneficial effects are as follows:

[0062] (1) Reduce model complexity: Traditional MTN needs to use high-order polynomials when dealing with highly nonlinear problems, which leads to complex models and large amounts of calculation. The present invention dynamically selects the most contributing features through a gradient scoring mechanism, effectively reducing model complexity, reducing the amount of calculation, and shortening the calculation time.

[0063] (2) Improving the systematicness of hyperparameter selection: The highest order of the MTN expansion term is used as a hyperparameter, and its selection is usually based on experiments, which is complex and lacks systematicity. The present invention provides a systematic method to select the optimal model order through gradient scoring, which reduces the dependence on experience and improves the adaptability and performance of the model in different data sets and application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a fault diagnosis neural network diagram of a dynamic multi-dimensional Taylor network based on gradient scoring of the present invention.

[0065] Figure 2 It is a framework diagram of the gradient-scoring-based dynamic multidimensional Taylor network (GSDMTN) used for fault diagnosis in the present invention. DETAILED DESCRIPTION

[0066] In order to better understand the technical content of the present invention, specific embodiments are given and described as follows in conjunction with the accompanying drawings.

[0067] Various aspects of the invention are described herein with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the invention are not limited to those described in the accompanying drawings. It should be understood that the invention is implemented by any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed in the invention are not limited to any implementation. In addition, some aspects disclosed in the invention may be used alone or in any appropriate combination with other aspects disclosed in the invention.

[0068] refer to Figure 2 The present invention provides a fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring, comprising the following steps:

[0069] S1. Based on the original signal of equipment fault, a fault signal feature dataset is constructed, including several original fault signals and corresponding fault labels;

[0070] S2. Construct a fault diagnosis neural network based on a dynamic multidimensional Taylor network with gradient scoring, train the fault diagnosis neural network with the original fault signal as input and the corresponding fault type as output, and obtain a fault diagnosis model; the fault diagnosis neural network includes an input layer, a first polynomial layer, a gradient scoring layer, a scoring sorting selection layer, a second polynomial layer, a first fully connected layer, a softmax layer, a second fully connected layer, a Sigmoid layer, and an output layer;

[0071] Fault-related variables are extracted from the original fault signal, and then the fault feature data is extracted. The input layer receives the fault feature data, and the fault feature data is generated and combined into different monomials through the first polynomial layer. The monomials are then input into the gradient scoring layer, and the automatic differentiation technology is used to calculate the gradient of the loss function relative to each monomial coefficient. The size of the gradient is used as a scoring standard to measure the contribution of each monomial to the approximation of the nonlinear function.

[0072] The scoring sorting selection layer is connected to the gradient scoring layer to sort the monomials according to the calculated gradient scores; and dynamically adjust the activation state of the monomials; and perform backpropagation based on the calculated gradients to update the parameters of each layer of the model, including the weights of the fully connected layer and the monomial coefficients of the polynomial layer.

[0073] On the other hand, the input layer is sequentially connected to the first fully connected layer and the Sigmoid layer, and the Sigmoid layer is connected to the scoring sorting selection layer. The monomials of the polynomial layer are screened according to the proportion of the nonlinear function represented by the selected monomials calculated from the gradient score, so that the monomials that meet the proportion requirements can participate in subsequent calculations.

[0074] The input end of the second polynomial layer is connected to the output end of the scoring sorting selection layer, the output end of the second polynomial layer is connected to the input end of the second fully connected layer, and the output end of the second fully connected layer is connected to the input end of the output layer of the Softmax layer to obtain the probability of each type. The Softmax layer is connected to the output layer to output the final classification result.

[0075] During the training process, the model has the ability to flexibly adjust the model structure according to the local characteristics or global structure of the input data. Specifically, it dynamically adjusts the activation state of the monomial and changes the combination of monomials involved in the calculation in the polynomial layer to adapt to different pattern complexities.

[0076] S3. Collect the original signal of equipment failure, input it into the fault diagnosis model, and obtain the fault classification result.

[0077] The network structure of the present invention refers to Figure 2 , the first polynomial layer receives the input vector x = (x1, x2, ..., x m), where m is the dimension of the input, and the output z is generated by combining different monomials, which include various power combinations of the input variables and cross terms. For example, if the inputs are x1 and x2, then the second-order terms may be and x1x2, etc. The first polynomial layer is a polynomial representation composed of monomials of different powers, which is used to approximate nonlinear functions; the output z of the first polynomial layer is (z1, z2, ..., z k ), k represents the dimension of the polynomial layer output, where each element z j The calculation formula is:

[0078]

[0079] in, Represents the monomial corresponding to the j-th output dimension The coefficient of n represents the number of input variables in the monomial. l ,l=1,2...,m represents the power of the corresponding input variable.

[0080] For example, if the polynomial layer has a second-order output term z1 for the input, it might contain and x1x2, then it can be expressed as:

[0081]

[0082] in and are the coefficients of the corresponding monomials.

[0083] The gradient scoring layer is connected to the polynomial layer and is used to calculate the gradient of the loss function with respect to each monomial coefficient using automatic differentiation techniques; the gradient is the rate of change of the loss function with respect to a certain variable (here, the monomial coefficient). In gradient-based optimization algorithms, the goal is to minimize the loss function. If the gradient of the loss function with respect to a certain monomial coefficient is large, it means that when we make a small change to the monomial coefficient, the loss function will change significantly. From the perspective of model optimization, those monomials that have a greater impact on the loss function (i.e., monomials with large gradients) play a more critical role in determining the output and performance of the model, so they can be considered more important. Compared with those monomials with smaller gradients, monomials with large gradients have a more significant impact on the loss function during model training. By calculating these gradients, the importance of each monomial in the process of the model approximating nonlinear functions can be quantified, thereby providing a basis for the subsequent dynamic adjustment of the activation state of the monomial.

[0084] Assuming there are K categories, the output of the model is The true label is y=(y1,y2,...,y K), where is a one-hot encoded vector that represents the category to which the sample belongs (only one element is 1, the rest are 0). The gradient layer is configured to perform the following steps:

[0085] S2A.1. Define the cross entropy loss function:

[0086]

[0087] in, is the output of the model, and the true label is y=(y1,y2,...,y K ), there are K fault categories.

[0088] S2A.2. Calculate the loss function L relative to the model output Gradient

[0089] According to the derivation formula get:

[0090]

[0091] When i=j,

[0092]

[0093] When i≠j,

[0094]

[0095] Right now

[0096]

[0097] S2A.3. Computational Model Output Relative to the polynomial output z j Gradient Model Output Where f is the Softmax activation function, then for the Softmax activation function,

[0098]

[0099] like Figure 2 As shown, the output of the second fully connected layer is connected to the input of the Softmax layer output layer to obtain the probability of each type, specifically:

[0100] The output of the second fully connected layer is input to the softmax layer. The output of the softmax layer is P = (p1, p2, ..., p K ), where p k represents the probability of belonging to the kth class, and the calculation formula is:

[0101]

[0102] Through the Softmax function, the output of the fully connected layer is converted into a probability distribution of each type.

[0103] The Softmax layer is connected to the output layer to output the final classification results, which are:

[0104] The output of the output layer is finally the classification result y, and its formula is:

[0105]

[0106] The monomials of the first polynomial layer and the second monomial layer include but are not limited to 1st order, 2nd order, 3rd order and higher order monomials, and the coefficient of each monomial is an adjustable parameter.

[0107] Relative to the monomial coefficient The gradient is as follows: The second polynomial output is transformed to obtain the model output Let the monomial of the polynomial be where x i is the input variable, n k is the power of the corresponding variable, m represents the number of variables in the monomial, and its coefficient is And the polynomial output is

[0108]

[0109] According to the polynomial derivation rule,

[0110]

[0111] According to the chain rule, the loss function L is relative to the monomial coefficient The gradient of is:

[0112]

[0113] According to the polynomial derivation rule, we can calculate:

[0114]

[0115] The number of monomials is M, let is the coefficient of the lth monomial, then

[0116]

[0117] The input layer is connected to the first fully connected layer and the Sigmoid layer in sequence, and the Sigmoid layer is connected to the scoring sorting selection layer. Let the input be x = (x1, x2, ..., x m), the weight matrix of the fully connected layer is Let the bias vector be 0, then the output of the first fully connected layer is calculated as:

[0118] s=Ω T x

[0119] The Sigmoid layer is connected to the fully connected layer and is used to screen the monomials of the polynomial layer according to the proportion of the selected monomials representing the nonlinear function calculated from the gradient score;

[0120] Assume that there are M monomials in the polynomial layer (the order is not distinguished here, and they are uniformly numbered). After the scoring and sorting selection layer is sorted according to the gradient score, the selection probability corresponding to the i-th monomial is p i (0≤p i ≤1), this probability is obtained through the Sigmoid function. The output s of the first fully connected layer is the output obtained through the Sigmoid function Each element p i ,i=1,2,...,m is calculated as follows:

[0121]

[0122] The values ​​in these outputs are between 0 and 1 and serve as ratios for selecting the monomials in the polynomial.

[0123] The scoring sorting selection layer is connected to the gradient scoring layer to sort the monomials according to the calculated gradient scores, which specifically includes the following sub-steps:

[0124] S2B.1. Suppose the polynomial layer has M monomials, whose coefficients are w in , i represents the monomial index, n represents the order-related identifier of the monomial, and the loss function calculated at the gradient level is relative to the i-th monomial coefficient w in The gradient is g i ,

[0125] S2B.2, Sorting operation: Define a sorting function S, which sorts the gradient vector g = (g1, g2, ..., g M ) to get the sorted index vector k=(k1,k2,…,k M ), so that g k1 ≥g k2 ≥…≥g kM , that is, k = S(g);

[0126] S2B.3. Select monomials based on the sorting results: Let the selected proportion be p i (0 <p i ≤1), then the number of monomials selected is

[0127] S2B.3. Selected monomial index k selected =(k1,k2,…,k N ), that is, the scoring sorting selection layer selects the monomial index according to the gradient score, and the subsequent model performs further calculations based on the selected monomial index.

[0128] The output of the second fully connected layer is connected to the input of the Softmax layer output layer to obtain the probability of each type, specifically:

[0129] The output of the second fully connected layer is input to the softmax layer. The output of the softmax layer is P = (p1, p2, ..., p K ), where p k represents the probability of belonging to the kth class, and the calculation formula is:

[0130]

[0131] Through the Softmax function, the output of the fully connected layer is converted into a probability distribution of each type.

[0132] The Softmax layer is connected to the fully connected layer to obtain the probability of each type;

[0133] The output of the fully connected layer is input to the softmax layer. The output of the softmax layer is P = (p1, p2, ..., p K ), where p k represents the probability of belonging to the kth class, and the calculation formula is:

[0134]

[0135] Through the Softmax function, the output of the fully connected layer is converted into a probability distribution of each type.

[0136] The output layer is connected to the Softmax layer to output the final classification results;

[0137] The output of the output layer is finally the classification result y, and its formula is:

[0138]

[0139] Where k = 1, 2, ..., K. That is, the output layer selects the category index with the largest vector probability in the probability vector output by the Softmax layer as the final classification result.

[0140] Although the present invention has been described above with preferred embodiments, it is not intended to limit the present invention. A person skilled in the art of the present invention may make various modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the definition of the claims.

Claims

1. A fault diagnosis method for a dynamic multidimensional Taylor network based on gradient scoring, characterized in that: The following steps are involved: S1. Based on the original signal of equipment fault, a fault signal feature dataset is constructed, including several original fault signals and corresponding fault labels; S2. Construct a fault diagnosis neural network based on a dynamic multidimensional Taylor network with gradient scoring, train the fault diagnosis neural network with the original fault signal as input and the corresponding fault type as output, and obtain a fault diagnosis model; the fault diagnosis neural network includes an input layer, a first polynomial layer, a gradient scoring layer, a scoring sorting selection layer, a second polynomial layer, a first fully connected layer, a softmax layer, a second fully connected layer, a Sigmoid layer, and an output layer; Extract fault-related variables from the original fault signal, and then extract fault feature data. The input layer receives the fault feature data, generates and combines the fault feature data into different monomials through the first polynomial layer, and then inputs the monomials into the gradient scoring layer. The gradient of the loss function relative to each monomial coefficient is calculated using automatic differentiation technology. The scoring and sorting selection layer is connected to the gradient scoring layer to sort the monomials according to the calculated gradient scores. On the other hand, the input layer is sequentially connected to the first fully connected layer and the Sigmoid layer, and the Sigmoid layer is connected to the scoring and sorting selection layer. The monomials of the polynomial layer are screened according to the proportion of the selected monomials calculated from the gradient scores that represent the nonlinear function. The input end of the second polynomial layer is connected to the output end of the scoring and sorting selection layer, the output end of the second polynomial layer is connected to the input end of the second fully connected layer, and the output end of the second fully connected layer is connected to the input end of the output layer of the Softmax layer to obtain the probability of each type. The Softmax layer is connected to the output layer to output the final classification result; S3. Collect the original signal of equipment failure, input it into the fault diagnosis model, and obtain the fault classification result.

2. The fault diagnosis method of the dynamic multidimensional Taylor network based on gradient scoring according to claim 1 is characterized in that: In step S2, the first polynomial layer receives an input vector x=(x1, x2, ..., x m ), where m is the dimension of the input, and the output z is generated by combining different monomials, where the monomials include various power combinations of the input variables and cross terms; The first polynomial layer is a polynomial representation composed of monomials of different powers of input, which is used to approximate nonlinear functions; The output of the first polynomial layer z=(z1,z2,...,z k ), k represents the dimension of the polynomial layer output, where each element z j The calculation formula is: in, Represents the monomial corresponding to the j-th output dimension The coefficient of n represents the number of input variables in the monomial. l ,l=1,2...,m represents the power of the corresponding input variable.

3. A fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring according to claim 2, characterized in that: The gradient layer is configured to perform the following steps: S2A.

1. Define the cross entropy loss function: in, is the output of the model, and the true label is y=(y1,y2,...,y K ), there are K fault categories. S2A.

2. Calculate the loss function L relative to the model output Gradient According to the derivation formula get: When i=j, When i≠j, Right now S2A.

3. Computational Model Output Relative to the polynomial output z j Gradient Model Output Where f is the Softmax activation function, then for the Softmax activation function, 4. The fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring according to claim 2 is characterized in that: Relative to the monomial coefficient The gradient is as follows: The second polynomial output is transformed to obtain the model output Let the monomial of the polynomial be where x i is the input variable, n k is the power of the corresponding variable, m represents the number of variables in the monomial, and its coefficient is And the polynomial output is According to the polynomial derivation rule, According to the chain rule, the loss function L is relative to the monomial coefficient The gradient of is: According to the polynomial derivation rule, we can calculate: The number of monomials is M, let is the coefficient of the lth monomial, then 5. The fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring according to claim 3 is characterized in that: The input layer is connected to the first fully connected layer and the Sigmoid layer in sequence, and the Sigmoid layer is connected to the scoring sorting selection layer. Let the input be x = (x1, x2, ..., x m ), the weight matrix of the fully connected layer is Let the bias vector be 0, then the output of the first fully connected layer is calculated as: s=Ω T x The polynomial layer has M monomials. After the scoring and sorting selection layer is sorted according to the gradient score, the selection probability corresponding to the i-th monomial is p i , 0≤p i ≤1, the probability is obtained through the Sigmoid function, The output s of the first fully connected layer is the output obtained by the Sigmoid function The calculation formula for each element pi, i = 1, 2, ..., m is: The values ​​in these outputs are between 0 and 1 and serve as ratios for selecting the monomials in the polynomial.

6. A fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring according to claim 5, characterized in that: The scoring sorting selection layer is connected to the gradient scoring layer to sort the monomials according to the calculated gradient scores, which specifically includes the following sub-steps: S2B.

1. Suppose the polynomial layer has M monomials, whose coefficients are w in , i represents the monomial index, n represents the order-related identifier of the monomial, and the loss function calculated at the gradient level is relative to the i-th monomial coefficient w in The gradient is g i , S2B.2, sorting operation: define a sorting function S, which sorts the gradient vector g = (g1, g2, ..., g M ) to get the sorted index vector k=(k1,k2,…,k M ), so that g k1 ≥g k2 ≥…≥g kM , that is, k = S(g); S2B.

3. Select monomials based on the sorting results: Let the selected proportion be p i (0 <p i ≤1), then the number of monomials selected is S2B.

3. Selected monomial index k selected =(k1,k2,…,k N ), that is, the scoring sorting selection layer selects the monomial index according to the gradient score, and the subsequent model performs further calculations based on the selected monomial index.

7. A fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring according to claim 6, characterized in that: The output of the second fully connected layer is connected to the input of the Softmax layer output layer to obtain the probability of each type, specifically: The output of the second fully connected layer is input to the softmax layer. The output of the softmax layer is P = (p1, p2, ..., p K ), where p k represents the probability of belonging to the kth class, and the calculation formula is: Through the Softmax function, the output of the fully connected layer is converted into a probability distribution of each type.

8. The fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring according to claim 7, characterized in that: The Softmax layer is connected to the output layer to output the final classification results, which are: The output of the output layer is finally the classification result y, and its formula is: Where k = 1, 2, ..., K. That is, the output layer selects the category index with the largest vector probability in the probability vector output by the Softmax layer as the final classification result.

9. The fault diagnosis method of a dynamic multidimensional Taylor network based on gradient scoring according to claim 7, characterized in that: The monomials of the first polynomial layer and the second monomial layer include but are not limited to 1st order, 2nd order, 3rd order and higher order monomials, and the coefficient of each monomial is an adjustable parameter.

Images