Chemical process fault diagnosis method based on EEMD decomposition and lightweight neural network

Abstract

The present invention relates to the field of chemical fault processing, and proposes a chemical process fault diagnosis method based on EEMD decomposition and lightweight neural network, which comprises three parts: a data decomposition method EEMD, a fault detection model AOS-LSTM, and a final model pruning. The specific steps of the chemical fault diagnosis method are: using the data decomposition method EEMD to decompose and preprocess the collected chemical fault data, and using the fault detection model AOS-LSTM to construct a backbone classification network. The LSTM is improved using an optimization algorithm, and the resulting AOS-LSTM performs better in classification performance than the original LSTM. Finally, the network model is compressed and pruned to make the model structure efficient and simple, which is conducive to model migration. Compared with the prior art, the present invention can quickly and conveniently diagnose faults generated in chemical processes.

Description

Technical Field

[0001] The present invention relates to the field of chemical process fault processing, and in particular to a chemical process fault diagnosis method based on EEMD decomposition and lightweight neural network. Background Art

[0002] Chemical engineering is a highly specialized and complex discipline, involving numerous fundamental theories and the application of new technologies. Numerous faults arise during chemical production processes within this field. Improving the efficiency of detecting these faults is a question worthy of in-depth research. Chemical process data exhibits nonlinear, high-dimensional, and non-Gaussian distributions, making the extraction of fault information more complex. Numerous traditional chemical process fault detection methods based on process data have been developed. These data-based methods do not require extensive prior expert knowledge; instead, they simply acquire the collected chemical process data and, by establishing appropriate fault detection models, can predict the current system status. Consequently, these methods are widely used in both scientific research and industrial applications. While popular traditional methods such as PCA, ICA, KPCA, KICA, and MICA can effectively detect some faults, their detection rates for certain perturbation faults are extremely low, indicating that these traditional methods still fail to fully and accurately extract information about these faults. This necessitates the development of new methods to improve fault detection rates. The nonlinearity, high noise, and non-Gaussian distribution characteristics of complex chemical processes mean that traditional chemical process fault detection methods do not demonstrate excellent diagnostic performance. Therefore, developing fault monitoring methods suitable for complex, nonlinear chemical processes is crucial. The layer-by-layer learning model of deep neural networks can more accurately learn the features implicit in the raw chemical process data, enabling the application of appropriate classification models to classify chemical process faults.

[0003] In recent years, with the continuous improvement and development of deep learning theory, the concept of time series has been introduced. This concept not only focuses on processing information at the current moment, but also connects information from previous and subsequent moments, thereby improving analysis accuracy. For example, RNNs incorporate this concept into network construction. Their essential feature is the integration of both internal feedback and feedforward connections between processing units within a network layer, enabling them to perceive information from multiple perspectives. However, RNNs have structural flaws that prevent them from effectively connecting long-range information, leading to vanishing and exploding gradients and making it impossible for the network to adjust for errors. Summary of the Invention

[0004] Purpose of the invention: To address the problem that faults generated in chemical processes cannot be diagnosed quickly and intelligently, the present invention provides a chemical process fault diagnosis method based on EEMD decomposition and lightweight neural network, which can diagnose faults generated in chemical processes conveniently and quickly.

[0005] Technical solution: The present invention discloses a chemical process fault diagnosis method based on EEMD decomposition and lightweight neural network, comprising the following steps:

[0006] Step 1: Obtain historical chemical industry time series data and use it as a training sample set and a test sample set; the training sample set is used to establish the intelligent fault detection model of this method, and the test sample set is used to verify the diagnostic accuracy of this method;

[0007] Step 2: Preprocess the acquired chemical industry historical time series data and use EEMD decomposition technology to decompose the original chemical industry data sequence;

[0008] Step 3: Build a network model and perform model training. Use the atomic orbital search algorithm (AOS) to optimize the three parameters of the LSTM model: the number of training times, the learning rate, and the number of hidden layer nodes. Use the fault detection model AOS-LSTM to build the backbone classification network.

[0009] Step 4: Compress and prune the network model in step 3. Use the BN layer in the network model in step 3 and add L1 to the scaling factor of the BN layer by pruning.

[0010] Step 5: Use the optimized network model to realize chemical process fault diagnosis.

[0011] Furthermore, the pre-processing in step 2 includes the following steps:

[0012] Step 2.1: For the acquired chemical time series data, use EEMD decomposition technology to decompose the original chemical data sequence, and add random Gaussian white noise x to the original signal data x(t). m (t) = x(t) + kn m (t) where k is the added noise amplitude coefficient;

[0013] Step 2.2: Then use the EMD method to analyze the signal x m (t) Decompose and obtain IMF components;

[0014] Step 2.3: Repeat steps 2.1 and 2.2, adding a different Gaussian white noise sequence each time;

[0015] Step 2.4: Record the mean of each IMF component obtained from the decomposition as the final decomposition result N is the number of times noise is added; c i，m represents the i-th IMF component obtained by the m-th EMD decomposition;

[0016] Step 2.5: Standardize the decomposed data and calculate the mean and standard deviation of the data. After the standardized data is obtained, μ is the mean of the sample data, σ is the standard deviation of the sample data, and after the standardization process, the training samples and test samples are appropriately selected.

[0017] Furthermore, the step 2.2 uses the EMD method to analyze the signal x m (t) is decomposed into:

[0018] First, find all the extreme points of the signal x(t), use a cubic spline curve to fit the envelope of the upper and lower extreme points, and calculate the average value m(t) of the upper and lower envelopes. Subtract m(t) from x(t) to get h(t). The number of extreme points and the number of zero-crossing points of the intermediate signal h(t) must be equal or differ by no more than one, and at any time, the average value of the upper envelope formed by the local maximum points and the lower envelope formed by the local minimum points is zero; if the conditions are met, h(t) is an IMF component. If not, replace x(t) with h(t) and repeat the above steps until h(t) meets the conditions to obtain the IMF component.

[0019] Furthermore, the specific operation of the atomic orbital search algorithm AOS for parameter optimization in step 3 is:

[0020] Step 3.1: First, initialize the candidate solution using the following formula to determine the initial position of the electron:

[0021]

[0022] in, represents the initial position of the candidate solution, are the minimum and maximum bounds of the jth decision variable of the i-th candidate solution, and rand is a uniformly distributed random vector in the range [0, 1];

[0023] Step 3.2: Set the number of electrons, the number of virtual layers in the region, and the dimension of the candidate solution; search for the learning rate, number of training times, and number of hidden layer nodes;

[0024] Use the cross entropy value as the electronic fitness function, calculate the electronic fitness value, update the optimal fitness value, and thus find the optimal model parameters. The fitness function is as follows:

[0025]

[0026] Among them, y is the actual output value, Output the predicted value; use the loss function value obtained by training LSTM as the fitness value of AOS. If the fitness value is better than the local optimal value and the global optimal value, update the values of the two. If the minimum error is met or the maximum number of iterations is reached, output the model parameters and use the obtained parameters to train the LSTM model. If not, update the algorithm.

[0027] Step 3.3: First calculate the position and fitness mean of the feasible solution of the K-th virtual layer, obtain the local optimal solution, and obtain the combined state BS of the k-th virtual layer k and binding energy BE k , the formula is as follows:

[0028]

[0029]

[0030] Step 3.4: Obtain the overall optimal solution by taking the average of the positions and fitness of all feasible solutions, and determine the overall binding state BS and binding energy BE formulas as follows:

[0031]

[0032]

[0033] Step 3.5: Set the number of iterations of the algorithm to 100. First, set a photon rate PR and generate a uniformly distributed random number φ for each electron in the range of (0, 1). When φ>PR and when The electron will move to the lower virtual layer, which updates the position of the electron. The update method is as follows:

[0034]

[0035] when The electrons will move to the higher virtual layer and the update method is as follows:

[0036]

[0037] α, β, and γ are all vectors of randomly generated numbers, uniformly distributed between (0, 1);

[0038] LE is the lowest energy level electron in the search domain, LE k is the electron in the lowest energy level in the kth virtual layer.

[0039] When φ<PR, the electron is affected by other effects instead of photons, and the electron position is updated as follows:

[0040]

[0041] where r is the vector of φ.

[0042] Furthermore, the specific operation of the LSTM model in step 4 is:

[0043] When the atomic orbital search algorithm AOS completes the parameter optimization iteration and reaches the end condition, the global optimal solution is output, and the optimal learning rate, training times and number of hidden layer nodes are used as the solution to determine the optimal parameters of LSTM. The parameters are used to build the AOS-LSTM model. LSTM includes memory cells c t And three gates: input gate i t , forget gate f t and output gate o t , when the input sequence is x t When , the outputs of the three thresholds and memory cells are as follows:

[0044] f t =σ(w f [h t-1 , x t ]+b f )

[0045] o t =σ(w o [h t-1 , x t ]+b o )

[0046] i t =σ(w i [h t-1 , x t ]+b i )

[0047] c t =f t ×c t-1 +i t ×tanh(w g [h t-1 , x t ]+b g )

[0048] In the above formula, σ is the sigmoid function, h t-1 is the output of the unit at the previous moment, w i 、w f 、w o 、w g is the model parameter, b f 、b i 、b o 、b g is the bias term;

[0049] After determining the optimal parameters of the LSTM network, the data was input into the LSTM network for training to extract feature information with time characteristics. A BN layer was added to the network, and the Adam optimizer was used to optimize the network. Finally, the Softmax classifier was used to classify the fault category.

[0050] Furthermore, the specific operation of pruning the network model in step 4 is:

[0051] The pruning method is used to add L1 to the scaling factor of the BN layer. L1 regularization performs sparse and feature selection on the weight parameters. The formula of L1 regularization is: min∑(y i -w T x i ) 2 +λ||w||1, where the first term is the loss function and the second term indicates that some values with smaller weights will be regularized;

[0052] Finally, the BN layer will perform the following conversion: z out =αz+β, where μ β and σ β are the mean and standard deviation of the input activations on the BN layer, α and β represent the scale and displacement, which provide the possibility of linearly transforming the normalized activations back to any scale, using the α parameter in the BN layer as the required scaling factor.

[0053] Beneficial effects:

[0054] This paper uses the EEMD method to decompose and preprocess collected chemical fault data. It then uses the AOS-LSTM fault detection model to construct a backbone classification network. By improving the LSTM using an optimization algorithm, the resulting AOS-LSTM achieves superior classification performance compared to the original LSTM. Finally, the network model is compressed and pruned, resulting in a more efficient and simplified model structure, facilitating model migration. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Flowchart of chemical fault detection method;

[0056] Figure 2 Modeling diagram of chemical process fault diagnosis model. Specific implementation plan

[0057] The present invention will be described in further detail below with reference to the examples, but the embodiments of the present invention are not limited thereto.

[0058] This paper discloses a chemical process fault diagnosis method based on EEMD decomposition and lightweight neural networks. The proposed method is applied to the Tennessee-Eastman (TE) benchmark chemical process for experimental purposes. The method specifically includes the following steps:

[0059] Step 1: Introduction to Chemical Data

[0060] The TE process primarily consists of five operating units: the reactor, condenser, vapor-liquid separator, recycle compressor, and stripping column. The simulated data includes 41 measured variables and 11 controlled variables. The TE process also includes 21 pre-defined faults. A description of the specific fault types is shown in Table 1. The sampling interval for both the training and test sets is 3 minutes. For each fault, the training set consists of 480 samples, and the test set consists of 960 samples.

[0061] Table 1 Pre-set faults in the TE process

[0062]

[0063]

[0064] Step 2: Preprocessing of chemical data

[0065] Step 2.1: First, perform EEMD decomposition on the data and add random Gaussian white noise x to the original signal data x(t) m (t) = x(t) + kn m (t) where k is the added noise amplitude coefficient.

[0066] Step 2.2: Then use the EMD method to analyze the signal x m To decompose the signal x(t), first find all the extreme points of the signal x(t). Use a cubic spline curve to fit the envelope of the upper and lower extreme points, and calculate the average value m(t) of the upper and lower envelopes. Subtract m(t) from x(t) to obtain h(t). The number of extreme points and the number of zero crossings of the intermediate signal h(t) must be equal or differ by no more than one. At any time, the average value of the upper envelope formed by the local maximum points and the lower envelope formed by the local minimum points must be zero. If the conditions are met, h(t) is an IMF component. If not, replace x(t) with h(t) and repeat the above steps until h(t) meets the conditions, thus obtaining the IMF component.

[0067] Step 2.3: Repeat steps 2.1 and 2.2, adding a different Gaussian white noise sequence each time.

[0068] Step 2.4: Record the mean of each IMF component obtained from the decomposition as the final decomposition result N is the number of times noise is added; c i，m represents the i-th IMF component obtained by the m-th EMD decomposition.

[0069] Step 2.5: Standardize the decomposed data and calculate the mean and standard deviation of the data. The standardized data is obtained, where μ is the mean of the sample data and σ is the standard deviation of the sample data. After standardization, training samples and test samples are obtained.

[0070] Step 3: Build a network model and train the model

[0071] First, the LSTM network parameters are optimized using the atomic orbital search algorithm to improve model accuracy. The AOS algorithm simulates the transition between two electron states, between high and low energy states. In this algorithm, each electron represents a feasible solution x, and the probability of an electron appearing in a layer in the search space follows a normal distribution.

[0072] Step 3.1: First, in the algorithm, the candidate solution is initialized by the following formula to determine the initial position of the electron:

[0073]

[0074] in, represents the initial position of the candidate solution, are the minimum and maximum bounds on the j-th decision variable of the ith candidate solution, and rand is a uniformly distributed random vector in the range [0, 1].

[0075] Step 3.2: Next, set the number of electrons to 30, the number of virtual layers in the region to 2, and the dimension of the candidate solution. Here, we search for the learning rate, the number of training cycles, and the number of hidden layer nodes, so the dimensionality is three-dimensional. The learning rate ranges from [0.001, 0.01], the number of training cycles ranges from [200, 1000], and the number of hidden layer nodes ranges from [10, 80].

[0076] Each electron has an energy state, which is the objective function value of the candidate solution. Here, the candidate solution with a better objective function value represents the electron with a lower energy state. The following vector equation is used to include the objective function values of different electrons:

[0077]

[0078] Where i is the number of electrons (feasible solutions), is the objective function value of the i-th solution in the k-th virtual layer.

[0079] Therefore, each candidate solution and the energy density corresponding to each candidate solution can be expressed as follows:

[0080]

[0081] in, is the i-th candidate solution in the k-th virtual layer, K is the k-th virtual layer, n is the maximum number of virtual layers; j is the dimension of the feasible solution.

[0082] The AOS algorithm calls the optimal fitness of the entire search domain LE, and the optimal solution in the kth virtual layer LE k , the electron position is optimized by pulling the global optimum and the current optimum. The lowest energy level electron (LE or LE k ) is the optimal candidate solution.

[0083] Here we use the cross entropy value as the electronic fitness function, calculate the electronic fitness value, update the optimal fitness value, and thus find the optimal model parameters. The fitness function is as follows:

[0084]

[0085] Among them, y is the actual output value, The predicted value is output. The loss function value obtained from LSTM training is used as the fitness value of AOS. If the fitness value is better than the local optimum and the global optimum, both values are updated. If the minimum error is met or the maximum number of iterations is reached, the model parameters are output and used to train the LSTM model. If not, the algorithm is updated.

[0086] Step 3.3: First calculate the position and fitness mean of the feasible solution of the K-th virtual layer to obtain the local optimal solution. Get the combined state BS of the k-th virtual layer k and binding energy BE k The formula is as follows:

[0087]

[0088]

[0089] Step 3.4: Then, the overall optimal solution is obtained by taking the average of the positions and fitness of all feasible solutions, and the overall binding state BS and binding energy BE formulas are determined as follows:

[0090]

[0091]

[0092] Step 3.5: Next, we need to update the algorithm. Here, we set the number of iterations to 100. First, we set a photon rate PR. Here, we set PR = 0.3. The algorithm will generate a uniformly distributed random number φ for each electron in the range of (0, 1). When φ>PR and when The electron will move to the lower virtual layer, which updates the position of the electron. The update method is as follows:

[0093]

[0094] when The electrons will move to the higher virtual layer and the update method is as follows:

[0095]

[0096] In the formula, α, β, and γ are all vectors of randomly generated numbers, uniformly distributed between (0, 1).

[0097] When φ<PR, the electron is affected by other effects instead of photons, and the electron position is updated as follows:

[0098]

[0099] Where r is the vector of φ.

[0100] Step 3.6: When the iteration is completed and the end condition is reached, the global optimal solution is output and the optimal learning rate, number of training times and number of hidden layer nodes are determined based on the solution. In other words, the optimal parameters of LSTM are determined and the AOS-LSTM model is constructed using the parameters. LSTM includes memory cells c t And three gates: input gate i t , forget gate f t and output gate o t , when the input sequence is x t When , the outputs of the three thresholds and memory cells are as follows:

[0101] f t =σ(w f [h t-1 , x t ]+b f )

[0102] o t =σ(w o [h t-1 , x t ]+b o )

[0103] i t =σ(w i [h t-1 , xt ]+b i )

[0104] c t =f t ×c t-1 +i t ×tanh(w g [h t-1 , x t ]+b g )

[0105] In the above formula, σ is the sigmoid function, h t-1 is the output of the unit at the previous moment, w i 、w f 、w o 、w g is the model parameter, b f 、b i 、b o 、b g is the bias term.

[0106] After determining the optimal parameters for the LSTM network, data was fed into the neural network for training, extracting temporal feature information. A batch normalization layer was also added to the network to prevent overfitting and facilitate convergence. The Adam optimizer was used to optimize the network. Finally, a Softmax classifier was used to classify the faults and obtain the final results.

[0107] Step 4: Prune the network model

[0108] The BN layer (batch normalization) is used in this model. This time, the pruning method is used to add L1 to the scaling factor of the BN layer. L1 regularization is used to prevent model overfitting and play a role in sparse model structure. L1 regularization performs sparseness and feature selection on the weight parameters.

[0109] The formula for L1 regularization is: min∑(y i -w T x i ) 2 +λ||w||1, where the first term is the loss function and the second term indicates that some values with smaller weights will be regularized.

[0110] Finally, the BN layer will perform the following conversion: z out =αz+β, where μ β and σ βis the mean and standard deviation of the input activation on the BN layer, α and β represent the scale and displacement, which provide the possibility of linearly transforming the normalized activation back to any scale. The α parameter in the BN layer is used as the required scaling factor.

[0111] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made in accordance with the spirit of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. A chemical process fault diagnosis method based on EEMD decomposition and lightweight neural network, characterized in that: The following steps are involved: Step 1: Obtain historical chemical industry time series data and use it as a training sample set and a test sample set; the training sample set is used to establish the intelligent fault detection model of this method, and the test sample set is used to verify the diagnostic accuracy of this method; Step 2: Preprocess the acquired chemical industry historical time series data and use EEMD decomposition technology to decompose the original chemical industry data sequence; Step 3: Build a network model and perform model training. Use the atomic orbital search algorithm (AOS) to optimize the three parameters of the LSTM model: the number of training times, the learning rate, and the number of hidden layer nodes. Use the fault detection model AOS-LSTM to build the backbone classification network. Step 3.1: First, initialize the candidate solution using the following formula to determine the initial position of the electron: in, represents the initial position of the candidate solution, are the minimum and maximum bounds of the jth decision variable of the i-th candidate solution, rand is a uniformly distributed random vector in the range [0,1]; Step 3.2: Set the number of electrons, the number of virtual layers in the region, and the dimension of the candidate solution; search for the learning rate, number of training times, and number of hidden layer nodes; Each electron has an energy state, which is the objective function value of the candidate solution. The candidate solution with a better objective function value represents the electron with a lower energy level. The following vector equation is used to include the objective function values of different electrons: Where i is the number of electrons, that is, the number of feasible solutions, is the objective function value of the i-th solution in the k-th virtual layer; Each candidate solution and the energy density corresponding to each candidate solution are expressed as follows: in, is the i-th candidate solution in the k-th virtual layer, k is the k-th virtual layer, n is the maximum number of virtual layers; j is the dimension of the feasible solution; The AOS algorithm calls the optimal fitness of the entire search domain LE, and the optimal solution in the kth virtual layer LE k , the electron position is optimized by pulling the global optimum and the current optimum, the lowest energy level electron LE or LE k is the optimal candidate solution; Use the cross entropy value as the electronic fitness function, calculate the electronic fitness value, update the optimal fitness value, and thus find the optimal model parameters. The fitness function is as follows: Among them, y is the actual output value, Output the predicted value; use the loss function value obtained by training LSTM as the fitness value of AOS. If the fitness value is better than the local optimal value and the global optimal value, update the values of the two. If the minimum error is met or the maximum number of iterations is reached, output the model parameters and use the obtained parameters to train the LSTM model. If not, update the algorithm. Step 3.3: First calculate the position and fitness mean of the feasible solution of the K-th virtual layer, obtain the local optimal solution, and obtain the combined state BS of the k-th virtual layer k and binding energy BE k , the formula is as follows: Step 3.4: Obtain the overall optimal solution by taking the average of the positions and fitness of all feasible solutions, and determine the overall binding state BS and binding energy BE formulas as follows: Step 3.5: Set the number of iterations of the algorithm to 100. First, set a photon rate PR and generate a uniformly distributed random number φ for each electron in the range (0,1). When φ>PR and when The electron will move to the lower virtual layer, which updates the position of the electron. The update method is as follows: when The electrons will move to the higher virtual layer and the update method is as follows: α, β, and γ are all vectors of randomly generated numbers, uniformly distributed between (0, 1); LE is the lowest energy level electron in the search domain, LE k is the electron with the lowest energy level in the kth virtual layer; When φ<PR, the electron is affected by other effects instead of photons, and the electron position is updated as follows: Where r is the vector of φ; Step 3.6: When the iteration is completed and the end condition is reached, the global optimal solution is output and the optimal learning rate, number of training times and number of hidden layer nodes are determined based on the solution. In other words, the optimal parameters of LSTM are determined and the AOS-LSTM model is constructed using the parameters. LSTM includes memory cells c t And three gates: input gate i t , forget gate f t and output gate o t , when the input sequence is x t When , the outputs of the three thresholds and memory cells are as follows: f t =σ(w f [h t-1 ,x t ]+b f ) the t =σ(w o [h t-1 ,x t ]+b o ) i t =σ(w i [h t-1 ,x t ]+b i ) c t =f t ×c t-1 +i t ×tanh(w g [h t-1 ,x t ]+b g ) In the above formula, σ is the sigmoid function, h t-1 is the output of the unit at the previous moment, w i 、w f 、w o 、w g is the model parameter, b f 、b i 、b o 、b g is the bias term; After determining the optimal parameters of the LSTM network, the data is input into the LSTM neural network for training. Feature information with time characteristics is extracted. A batch normalization layer is added to the network, and the Adam optimizer is used to optimize the network. Finally, the Softmax classifier is used to classify the fault categories to obtain the final results. Step 4: Compress and prune the network model in step 3. Use the BN layer in the network model in step 3 and add L1 to the scaling factor of the BN layer by pruning. Step 5: Use the optimized network model to realize chemical process fault diagnosis.

2. The chemical process fault diagnosis method based on EEMD decomposition and lightweight neural network according to claim 1 is characterized in that: The pre-processing in step 2 comprises the following steps: Step 2.1: For the acquired chemical time series data, use EEMD decomposition technology to decompose the original chemical data sequence, and add random Gaussian white noise x to the original signal data x(t). m (t) = x(t) + kn m (t), where k is the added noise amplitude coefficient; Step 2.2: Then use the EMD method to analyze the signal x m (t) Decompose and obtain IMF components; Step 2.3: Repeat steps 2.1 and 2.2, adding a different Gaussian white noise sequence each time; Step 2.4: Record the mean of each IMF component obtained from the decomposition as the final decomposition result N is the number of times noise is added; c i,m represents the i-th IMF component obtained by the m-th EMD decomposition; Step 2.5: Standardize the decomposed data and calculate the mean and standard deviation of the data. After the standardized data is obtained, μ is the mean of the sample data, σ is the standard deviation of the sample data, and after the standardization process, the training samples and test samples are appropriately selected.

3. The chemical process fault diagnosis method based on EEMD decomposition and lightweight neural network according to claim 2 is characterized in that: The step 2.2 uses the EMD method to analyze the signal x m (t) is decomposed into: First, find all the extreme points of the signal x(t), use a cubic spline curve to fit the envelope of the upper and lower extreme points, and calculate the average value m(t) of the upper and lower envelopes. Subtract m(t) from x(t) to get h(t). The number of extreme points and the number of zero-crossing points of the intermediate signal h(t) must be equal or differ by no more than one, and at any time, the average value of the upper envelope formed by the local maximum points and the lower envelope formed by the local minimum points is zero; if the conditions are met, h(t) is an IMF component. If not, replace x(t) with h(t) and repeat the above steps until h(t) meets the conditions to obtain the IMF component.

4. The chemical process fault diagnosis method based on EEMD decomposition and lightweight neural network according to claim 1 is characterized in that: The specific operation of pruning the network model in step 4 is: The pruning method is used to add L1 to the scaling factor of the BN layer. L1 regularization performs sparse and feature selection on the weight parameters. The formula of L1 regularization is: min∑(y i -w T x i ) 2 +λ||w||1, where the first term is the loss function and the second term indicates that some values with smaller weights will be regularized; Finally, the BN layer will perform the following conversion: z out =αz+β, where μ β and σ β are the mean and standard deviation of the input activations on the BN layer, α and β represent the scale and displacement, which provide the possibility of linearly transforming the normalized activations back to any scale, using the α parameter in the BN layer as the required scaling factor.

Images