A Smart BIT Design Method Based on Improved Deep Forest

Abstract

This invention relates to the field of artificial intelligence, and more particularly to an intelligent BIT design method based on an improved deep forest, comprising: S1, acquiring BIT data from the analog output module of a heavy-duty gas turbine control system to form a historical dataset of functional circuit BIT sampling for the analog output module; S2, extracting features from the historical dataset of functional circuit BIT sampling using an empirical mode decomposition algorithm; S3, training a random forest decision tree using a classification tree regression algorithm on the feature vectors extracted in S2, and training a rotated forest using a rotation strategy on the extracted feature vectors; S4, training the improved deep forest, acquiring an online dataset of functional circuit BIT sampling for the analog output module, and identifying three states—normal, intermittent fault, and permanent fault—of the online dataset of functional circuit BIT sampling using the improved deep forest; this invention achieves accurate classification of the BIT state of the analog output channel circuit by constructing a deep forest.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more particularly to an intelligent BIT design method based on an improved deep forest. Background Technology

[0002] In heavy-duty gas turbine control systems, the controller can transmit analog signals to field actuators such as valves and feedwater pumps via analog output modules, enabling long-distance control. The reliability of the analog output modules significantly impacts the stability of the entire control task, ultimately determining its success or failure. To improve the reliability of analog output modules, many researchers have incorporated Built-in Test (BIT) technology, designing BIT detection circuits using closed-loop loop testing methods. However, conventional BIT technology often suffers from a high false alarm rate. This high false alarm rate severely negatively impacts the reliability of the analog output modules, leading to a loss of user confidence in the entire control system. Furthermore, a high false alarm rate makes the re-inspection and maintenance of the analog output modules difficult.

[0003] Analog output modules typically include a D / A output, signal amplification circuitry, and a V / I (voltage / current) conversion circuit. In practical operation, analog output modules frequently experience intermittent faults due to factors such as open negative feedback in operational amplifiers, aging resistors, open diodes, and short circuits in transistors. However, because intermittent fault signals are random and lack a clear frequency or waveform, while also containing characteristics of both permanent fault signals and normal signals, the sample set of intermittent fault signals collected in actual operation is relatively small, accounting for only a fraction of a percent of the total dataset. Therefore, intermittent fault signals are characterized by high acquisition costs, difficulty in identification, recurrence, and small sample sets. Consequently, effectively reducing the false alarm rate (BIT) of analog output modules and improving their reliability has become a pressing issue that needs to be addressed. Summary of the Invention

[0004] To address this, the present invention provides an intelligent BIT design method based on improved deep forest, which overcomes the problems of high false alarm rate and low reliability of analog output module BIT in the prior art.

[0005] To achieve the above objectives, the present invention provides an intelligent BIT design method based on an improved deep forest, comprising the following steps: S1, acquiring BIT data from the analog output module of a heavy-duty gas turbine control system to form a functional circuit BIT sampling history dataset of the analog output module.

[0006] Step S2: Use the empirical mode decomposition algorithm to extract features from the historical BIT sampling dataset of the functional circuit.

[0007] Step S3: Use the feature vectors extracted in step S2 to train a random forest decision tree using a classification tree regression algorithm; and use the extracted feature vectors to train a rotation forest using a rotation strategy.

[0008] Step S4: Train the improved deep forest, collect the online dataset of functional circuit BIT sampling of the analog output module, and use the improved deep forest to identify the three states of normal, intermittent fault and permanent fault of the online dataset of functional circuit BIT sampling.

[0009] Furthermore, the current or voltage representing the functional circuit state of the analog output channel circuit under different operating conditions is sampled, including normal data of the main functional circuit of the analog output channel circuit, the permanent fault data, and the intermittent fault data, to obtain the functional circuit BIT sampling dataset. The empirical mode decomposition algorithm is used to extract features from the sampling dataset, and the feature vector for each state is constructed as input for training the decision tree, including the following steps.

[0010] The first step is to find all the extreme points in the dataset X, and then connect the local maxima into an upper envelope and the local minima into a lower envelope using spline curves. The area between the upper and lower envelopes contains all the data points.

[0011] The second step is to derive h1 = X - m1 from the average value m1 of the upper and lower envelopes. If h1 satisfies the intrinsic modal component, the number of extreme points and the number of zero crossings must be equal or differ by no more than 1 in the entire data segment. At any time, if the average value of the upper envelope formed by the local maxima and the lower envelope formed by the local minima is zero, then h1 can be considered as the first IMF component of X.

[0012] Third, if the IMF conditions are not met, h1 is used as the original data, and the first and second steps are repeated to obtain the mean m11 of the upper and lower envelopes. The condition for the IMF component is then calculated by checking whether h11 = h1 - m11 is suitable. If not, the above two steps are repeated p times until the IMF conditions are met, and the first IMF component c1 is obtained, where c1 = h1p.

[0013] Step 4: Separate c1 from X to obtain r1, r1 = X - c1. Use r1 as the original signal and repeat steps 1 to 3, looping q times, to obtain the second to the qth IMF components c2 to cq.

[0014]

[0015] Fifth step: When rq becomes a monotonic function, rq is called the residual component, and the decomposition stops. At this point, the original signal is decomposed into q IMF components and the sum of a single-point function rq, that is:

[0016]

[0017] Step 6: Construct feature vectors using each IMF component.

[0018]

[0019] After extracting EMD features from all samples, the dataset D is formed:

[0020]

[0021] Where T n The feature vector representing the nth state. This represents the q-th sample vector of the nth type.

[0022] Optionally, to ensure the uniformity and consistency of the data distribution, k-fold cross-validation is performed on the extracted feature vectors. The dataset D is composed of feature vectors T. n The dataset D is divided into k mutually exclusive subsets of equal size. Each time, the union of k-1 subsets is used as the training set, and the other subset is used as the validation set, resulting in k sets of training samples. These k sets of data are then used as the input for training the decision tree.

[0023] The k datasets are used to train a random forest decision tree using the CART algorithm. A subset containing A′ = log₂A attributes is randomly selected from all A attributes in dataset D. Then, the attribute with the smallest Gini index in A′ is selected as the next branch feature. The formula for the Gini index is:

[0024]

[0025] The growth direction of the random forest decision tree branches is determined by Gini(A). The split stops when all samples contained in the current node belong to the same category, or when the number of samples covered by the current node is less than a certain threshold. Then, the tree is pruned appropriately using the Reduced-Error Pruning (REP) method to produce a random forest decision tree. The remaining k-1 sets of data are trained using the CART algorithm to obtain k random forest decision trees.

[0026] The extracted feature vectors are used to train a rotation forest using a rotation strategy.

[0027] Further, each training set is resampled by 75% to obtain training samples Dk in the dataset D. Principal Component Analysis (PCA) is used to transform the training samples Dk to obtain the principal component coefficient eigenvectors Pk = [p1k, p2k, ..., pdk]. The k eigenvectors are then combined into a diagonal coefficient matrix R.

[0028]

[0029] The matrix R* is obtained by rotating the original sample feature positions corresponding to the feature vectors, and a new sample Xnew = X*R* is reconstructed. Xnew is then used to train a rotated forest decision tree.

[0030] The random forest and rotation forest are combined to form an improved deep forest layer, and the decision results of the random forest and rotation forest, along with the dataset D, are used as input to the next layer of the deep forest.

[0031] In step S4, the BIT detection data of the main functional circuit of the analog output module is used to extract feature vectors using the EMD algorithm, and then input into the improved deep forest network for voting selection to complete the state identification. The identified states include normal state, intermittent fault state and permanent fault state.

[0032] When making a vote, the voting process is as follows:

[0033]

[0034]

[0035] Here, hi represents different forests in the improved depth forest. For each forest, the predicted results, The size is used as the basis for voting, and the choice is made based on the size of the vote. The largest The final improved deep forest classification result is the identification result of the functional circuit state of the analog output module.

[0036] Compared with the prior art, the beneficial effect of the present invention is that, in response to the problem of high false alarm rate of BIT in analog output module of heavy-duty gas turbine control system, the present invention improves the deep forest algorithm to classify the state of BIT detection data of analog output module, accurately identifies the intermittent state of analog output module, effectively reduces the false alarm rate of analog output module BIT, and provides an effective method for improving reliability. Attached Figure Description

[0037] Figure 1This is a schematic diagram of the overall process of the intelligent BIT based on the improved deep forest described in this embodiment;

[0038] Figure 2 This is a flowchart of the intelligent BIT design based on improved deep forest described in this embodiment;

[0039] Figure 3 This is a flowchart of the k-fold cross-validation process described in this embodiment;

[0040] Figure 4 This is a flowchart of the improved deep forest decision-making process described in this embodiment. Detailed Implementation

[0041] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0042] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0043] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0044] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0045] Please see Figure 1 As shown, it is a schematic diagram of the overall process of the intelligent BIT based on the improved deep forest described in this embodiment. This embodiment discloses an intelligent BIT design method based on the improved deep forest, including step S1, collecting BIT data from the analog output module of the heavy-duty gas turbine control system to form a functional circuit BIT sampling history dataset of the analog output module.

[0046] Step S2: Use the empirical mode decomposition algorithm to extract features from the historical BIT sampling dataset of the functional circuit.

[0047] Step S3: Use the feature vectors extracted in step S2 to train a random forest decision tree using a classification tree regression algorithm; and use the extracted feature vectors to train a rotation forest using a rotation strategy.

[0048] Step S4: Train the improved deep forest, collect the online dataset of functional circuit BIT sampling of the analog output module, and use the improved deep forest to identify the three states of normal, intermittent fault and permanent fault of the online dataset of functional circuit BIT sampling.

[0049] Please continue reading. Figure 2 The diagram shown is a flowchart of the intelligent BIT design based on improved deep forest described in this embodiment. Specifically, the current or voltage representing the functional circuit state of the analog output channel circuit under different operating states is sampled, including normal data of the main functional circuit of the analog output channel circuit, permanent fault data, and intermittent fault data, to obtain the functional circuit BIT sampling dataset. The empirical mode decomposition algorithm is used to extract features from the sampling dataset, and the feature vector for each state is constructed as the input for training the decision tree. The process includes the following steps.

[0050] The first step is to find all the extreme points in the dataset X, and then connect the local maxima into an upper envelope and the local minima into a lower envelope using spline curves. The area between the upper and lower envelopes contains all the data points.

[0051] The second step is to derive h1 = X - m1 from the average value m1 of the upper and lower envelopes. If h1 satisfies the intrinsic modal component, the number of extreme points and the number of zero crossings must be equal or differ by no more than 1 in the entire data segment. At any time, if the average value of the upper envelope formed by the local maxima and the lower envelope formed by the local minima is zero, then h1 can be considered as the first IMF component of X.

[0052] Third, if the IMF conditions are not met, h1 is used as the original data, and the first and second steps are repeated to obtain the mean m11 of the upper and lower envelopes. The condition for the IMF component is then calculated by checking whether h11 = h1 - m11 is suitable. If not, the above two steps are repeated p times until the IMF conditions are met, and the first IMF component c1 is obtained, where c1 = h1p.

[0053] Step 4: Separate c1 from X to obtain r1, r1 = X - c1. Use r1 as the original signal and repeat steps 1 to 3, looping q times, to obtain the second to the qth IMF components c2 to cq.

[0054]

[0055] Fifth step: When rq becomes a monotonic function, rq is called the residual component, and the decomposition stops. At this point, the original signal is decomposed into q IMF components and the sum of a single-point function rq, that is:

[0056]

[0057] Step 6: Construct feature vectors using each IMF component.

[0058]

[0059] After extracting EMD features from all samples, the dataset D is formed:

[0060]

[0061] Where T n The feature vector representing the nth state. This represents the q-th sample vector of the nth type.

[0062] Please continue reading. Figure 3 As shown, this is a flowchart of the k-fold cross-validation process described in this embodiment. Specifically, to ensure the uniformity and consistency of the data distribution, k-fold cross-validation is performed on the extracted feature vectors. The dataset D is composed of feature vector T. n The dataset D is divided into k mutually exclusive subsets of equal size. Each time, the union of k-1 subsets is used as the training set, and the other subset is used as the validation set, resulting in k sets of training samples. These k sets of data are then used as the input for training the decision tree.

[0063] The k datasets are used to train a random forest decision tree using the CART algorithm. A subset containing A′ = log₂A attributes is randomly selected from all A attributes in dataset D. Then, the attribute with the smallest Gini index in A′ is selected as the next branch feature. The formula for the Gini index is:

[0064]

[0065] The growth direction of the random forest decision tree branches is determined by Gini(A). The split stops when all samples contained in the current node belong to the same category, or when the number of samples covered by the current node is less than a certain threshold. Then, the tree is pruned appropriately using the Reduced-Error Pruning (REP) method to produce a random forest decision tree. The remaining k-1 sets of data are trained using the CART algorithm to obtain k random forest decision trees.

[0066] The extracted feature vectors are used to train a rotation forest using a rotation strategy.

[0067] Further, each training set is resampled by 75% to obtain training samples Dk in the dataset D. Principal Component Analysis (PCA) is used to transform the training samples Dk to obtain the principal component coefficient eigenvectors Pk = [p1k, p2k, ..., pdk]. The k eigenvectors are then combined into a diagonal coefficient matrix R.

[0068]

[0069] The matrix R* is obtained by rotating the original sample feature positions corresponding to the feature vectors, and a new sample Xnew = X*R* is reconstructed. Xnew is then used to train a rotated forest decision tree.

[0070] Please continue reading. Figure 4 As shown, it is a flowchart of the improved deep forest decision-making process described in this embodiment. Specifically, random forest and rotation forest are combined to form one layer of the improved deep forest, and the decision results of random forest and rotation forest, along with the dataset D, are used as the input to the next layer of the deep forest.

[0071] In step S4, the BIT detection data of the main functional circuit of the analog output module is used to extract feature vectors using the EMD algorithm, and then input into the improved deep forest network for voting selection to complete the state identification. The identified states include normal state, intermittent fault state and permanent fault state.

[0072] When making a vote, the voting process is as follows:

[0073]

[0074]

[0075] Here, hi represents different forests in the improved depth forest. For each forest, the predicted results, The size is used as the basis for voting, and the choice is made based on the size of the vote. The largest The final improved deep forest classification result is the identification result of the functional circuit state of the analog output module.

[0076] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent BIT design method based on improved deep forest, characterized in that, Comprising, Step S1, BIT data acquisition is performed on the analog output module of the heavy gas turbine control system to form a functional circuit BIT sampling historical data set of the analog output module; Step S2, the empirical mode decomposition algorithm is used to extract features from the functional circuit BIT sampling historical data set; Step S3, the feature vectors extracted in step S2 are trained using a classification tree regression algorithm to train a random forest decision tree, and the extracted feature vectors are trained using a rotation strategy to train a rotated forest; Step S4, an improved deep forest is trained, a BIT sampling online data set of the functional circuit of the analog output module is collected, and the functional circuit BIT sampling online data set is identified as normal, intermittent fault and permanent fault by the improved deep forest; The current or voltage of the functional circuit of the analog output module under different working conditions can represent the state of the functional circuit, including normal data, permanent fault data and intermittent fault data of the functional circuit of the analog output module, to obtain the functional circuit BIT sampling historical data set, the empirical mode decomposition algorithm is used to extract features from the functional circuit BIT sampling historical data set, and the feature vectors of each state are constructed as the input of the decision tree, comprising, First step, find all extreme points in the functional circuit BIT sampling historical data set X, connect the local maximum points into an upper envelope line and the local minimum points into a lower envelope line by a spline curve, and all data points are contained between the upper and lower envelope lines; Second step, from the average value m1 of the upper and lower envelope lines, h1=X-m1, if h1 satisfies the intrinsic mode component, the number of extreme points and the number of zero-crossing points must be equal or differ by at most 1 in the entire data segment, and the average value of the upper envelope line formed by the local maximum points and the lower envelope line formed by the local minimum points is zero at any time, then h1 is considered as the first IMF component of X; Third step, if it does not meet the IMF condition, h1 is taken as the original data, and the first step and the second step are repeated to obtain the average values m11 of the upper and lower envelopes, and h11=h1-m11 is calculated to determine whether it meets the necessary condition of the IMF component, if not, repeat the above two steps p times until the IMF condition is met, and the first IMF component c1 is obtained, c1=h1p; Fourth step, separate c1 from X to obtain r1, r1=X-c1, and repeat the first step to the third step with r1 as the original signal for q times to obtain the second to qth IMF components c2 to cq, then Fifth step, when rq becomes a monotonic function, rq is called a residual component, and the decomposition is stopped, at this time the original signal is decomposed into the sum of q IMF components and a single point function rq, that is, Sixth step, the feature vectors are constructed using the IMF components, After EMD feature extraction is performed on all samples, a data set D is formed, Comprising, Step S1, BIT data acquisition is performed on the analog output module of the heavy gas turbine control system to form a functional circuit BIT sampling historical data set of the analog output module; Step S2, the empirical mode decomposition algorithm is used to extract features from the functional circuit BIT sampling historical data set; Step S3, the feature vectors extracted in step S2 are trained using a classification tree regression algorithm to train a random forest decision tree, and the extracted feature vectors are trained using a rotation strategy to train a rotated forest; Step S4, an improved deep forest is trained, a BIT sampling online data set of the functional circuit of the analog output module is collected, and the functional circuit BIT sampling online data set is identified as normal, intermittent fault and permanent fault by the improved deep forest; The current or voltage of the functional circuit of the analog output module under different working conditions can represent the state of the functional circuit, including normal data, permanent fault data and intermittent fault data of the functional circuit of the analog output module, to obtain the functional circuit BIT sampling historical data set, the empirical mode decomposition algorithm is used to extract features from the functional circuit BIT sampling historical data set, and the feature vectors of each state are constructed as the input of the decision tree, comprising, First step, find all extreme points in the functional circuit BIT sampling historical data set X, connect the local maximum points into an upper envelope line and the local minimum points into a lower envelope line by a spline curve, and all data points are contained between the upper and lower envelope lines; Second step, from the average value m1 of the upper and lower envelope lines, h1=X-m1, if h1 satisfies the intrinsic mode component, the number of extreme points and the number of zero-crossing points must be equal or differ by at most 1 in the entire data segment, and the average value of the upper envelope line formed by the local maximum points and the lower envelope line formed by the local minimum points is zero at any time, then h1 is considered as the first IMF component of X; Third step, if it does not meet the IMF condition, h1 is taken as the original data, and the first step and the second step are repeated to obtain the average values m11 of the upper and lower envelopes, and h11=h1-m11 is calculated to determine whether it meets the necessary condition of the IMF component, if not, repeat the above two steps p times until the IMF condition is met, and the first IMF component c1 is obtained, c1=h1p; Fourth step, separate c1 from X to obtain r1, r1=X-c1, and repeat the first step to the third step with r1 as the original signal for q times to obtain the second to qth IMF components c2 to cq, then Fifth step, when rq becomes a monotonic function, rq is called a residual component, and the decomposition is stopped, at this time the original signal is decomposed into the sum of q IMF components and a single point function rq, that is, Sixth step, the feature vectors are constructed using the IMF components, After EMD feature extraction is performed on all samples, a data set D is formed, where T n characteristic vector representing the nth state, qth IMF component representing the nth state; The extracted feature vectors are subjected to k-fold cross validation, and the data set D is composed of the feature vectors T n The data set D is equally divided into k mutually exclusive subsets, and each time the union of k-1 subsets is used as the training set and the other subset is used as the validation set, and finally k sets of training samples are obtained, which are used as the input of the training decision tree. The k groups of data sets are trained by using the CART algorithm to train the random forest decision tree, a subset containing A' = log2A attributes is randomly selected from all A attributes of the data set D, and the attribute feature with the minimum current data Gini index is selected as the attribute feature for the next branch, and the Gini(A) formula is used to determine the growth direction of the random forest decision tree branch. The samples contained in the current node belong to the same category, or the number of samples covered by the current node is less than a certain threshold value, and the splitting stops, and then the error rate reduction pruning method is used to prune the tree to produce a random forest decision tree. The remaining k-1 groups of data are trained by using the CART algorithm to train the random forest decision tree, and k random forest decision trees are obtained. The extracted feature vectors are trained by using the rotation strategy to train the rotation forest; Each group of training sets is resampled by 75% to obtain the training samples Dk in the data set D, principal component analysis is used to transform the training samples Dk to obtain the principal component coefficient feature vector Pk = [p1k, p2k, …, pdk] of the training samples, and the k feature vectors are combined into a diagonal coefficient matrix R, and the matrix R* is obtained by re-rotating according to the original sample feature position corresponding to the feature vector, and the new sample Xnew = X*R* is reconstructed, and the rotation forest decision tree is trained by using Xnew; The random forest and the rotation forest are combined to form a layer of the improved deep forest, and the decision results of the random forest and the rotation forest are combined to form a layer of the improved deep forest. In the step S4, after the BIT detection data of the functional circuit of the analog output module is extracted into a feature vector by using the EMD algorithm, the feature vector is input into the improved deep forest network, a vote is selected, and the state recognition is completed. The recognized states include a normal state, an intermittent fault state and a permanent fault state.

2. The improved deep forest based intelligent BIT design method of claim 1, wherein, In the voting selection, the voting process is 3. The improved deep forest based intelligent BIT design method of claim 2, wherein, ​ wherein hz represents different forests in the improved deep forest, the prediction results of each forest, the size as the basis for voting, select can make the largest as the final improved deep forest classification results, that is, the identification results of the analog output module function circuit state.

Images