Fault diagnosis method based on semi-supervised random configuration network
By utilizing unlabeled sample information and manifold regularization strategies through a semi-supervised randomized network, the problem of low fault diagnosis accuracy in complex industrial scenarios is solved, achieving fast and accurate fault identification and fault diagnosis with low hardware requirements.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUIZHOU UNIV
- Filing Date
- 2025-11-27
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies have low accuracy in fault diagnosis in complex industrial scenarios. Deep learning methods require a large amount of labeled data and hardware resources, while traditional methods are difficult to capture nonlinear coupling relationships and consume a lot of computational resources.
A semi-supervised randomized network is adopted, which utilizes unlabeled sample information and constructs a lightweight network structure through manifold regularization and joint optimization strategies. This avoids the backpropagation process, adaptively adjusts weights and biases, and designs an inequality constraint mechanism to improve recognition accuracy.
It enables fast and accurate fault identification in resource-limited environments, improves the algorithm's identification accuracy and training speed, and reduces hardware requirements.
Smart Images
Figure CN122019218A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault diagnosis technology, and specifically to a fault diagnosis method based on a semi-supervised random configuration network. Background Technology
[0002] Fault diagnosis is an important technology for large-scale industrial maintenance. With the popularization of automation technology and the widespread application of sensor technology, industrial equipment can collect a large amount of operating data. Making good use of this data can enable fault diagnosis to play a vital role in system safety and process reliability.
[0003] Currently, rule-based, model-based, and statistical methods are commonly used for industrial fault diagnosis, and they perform well in certain applications. However, in complex scenarios, all three methods are subject to limitations. Rule-based fault diagnosis methods struggle to uncover the deep correlation between fault categories and collected data using expert experience and knowledge. Model-based methods find it difficult to build complex physical models, and statistical fault diagnosis methods struggle to capture the nonlinear coupling between fault categories and collected data and have limited general applicability. Therefore, these three methods generally have low accuracy in fault diagnosis in complex industrial scenarios.
[0004] In recent years, deep learning has received considerable attention in the field of intelligent manufacturing. Through multi-layered neural networks, deep learning methods possess powerful feature learning capabilities and exhibit excellent generalization performance when dealing with high-dimensional nonlinear data. However, deep learning methods have limitations. They require a large amount of labeled data for training, but in real-world industrial fault diagnosis, the cost of labeling data is high, resulting in limited labeled data. Furthermore, deep learning methods require significant hardware resources during training and inference, limiting their application in resource-constrained environments. For example, CNN and transform methods require multiple convolutional and pooling layers, resulting in complex network structures and the need to solve gradients layer by layer through backpropagation. This makes training CNNs computationally expensive and time-consuming. This computational bottleneck is particularly prominent in industrial applications, exposing the inherent limitations of traditional deep architectures. To overcome these problems, academia is developing lightweight network structures with fast training speeds and fewer parameters for semi-supervised industrial fault diagnosis. Summary of the Invention
[0005] The technical problem solved by this invention is to provide a fault diagnosis method based on a semi-supervised random configuration network, and to provide a lightweight network structure with fast training and few parameters.
[0006] The basic solution provided by this invention is a fault diagnosis method based on a semi-supervised random configuration network, comprising the following steps: The method includes the following steps: S1. Vibration signal acquisition is performed using specialized equipment for key components of industrial equipment; S2. Data preprocessing: Analyze the collected signals and extract features; standardize the data to eliminate the influence of dimensions, thereby improving the model training effect; S3. Initialize the semi-supervised randomized network: Split the network into unlabeled samples, labeled samples, and a test set; set the Laplacian matrix parameters and construct the Laplacian matrix; determine the optimal number of hidden layer nodes. Critical tolerance error Number of candidate nodes Set the range of random parameters and ; S4. Randomly configure the weights and biases of the nodes and construct candidate neural nodes for the hidden layer; S5. Select candidate nodes using inequality constraints; S6. Determine the optimal neural node and add it to the network model; S7. Initialize the unlabeled sample indicator matrix; S8. Calculate the output weight matrix and use a joint optimization method to iteratively optimize the index matrix and output weight matrix of the unlabeled samples; S3 specifically includes: given a set of preprocessed data, inputting unlabeled samples, labeled samples, and a test set; initializing a semi-supervised randomized network. Input dataset ,in For sample features, For sample labels, ,in For labeled sample features, To indicate the number of tags, This represents the number of unlabeled samples. , For X L The tag, Each element in is equal to Therefore The training process is unknown. l + u = n Initialize the Laplace matrix parameters and determine the optimal number of hidden layer nodes. Critical tolerance error Number of candidate nodes ,set up The selection range is , r The selection range is Regularization coefficient and The selection range is as follows: The optimization objective of the semi-supervised randomized network is as follows:
[0007]
[0008]
[0009]
[0010] in h For activation function, This is the output of the Kth hidden node. For Laplace matrix, The calculation method is as follows:
[0011] in yes x of k A collection of fields The average edge length of the local adjacency graph; Here, the adjacency matrix S is derived from... The matrix D is a diagonal matrix formed by summing each column of the adjacency matrix S.
[0012] Furthermore, S4 specifically includes: from In the process, output weights are randomly selected. and bias From a sequence of nonnegative real numbers Select r Substitute into the activation function get Group candidate neural nodes .
[0013] Furthermore, S5 specifically includes: the semi-supervised randomized network is a single-hidden-layer network, and the network model is constructed by adding nodes one by one. Assuming that a network model containing K-1 hidden-layer nodes has already been constructed, the newly added Kth neuron node in the semi-supervised randomized network needs to satisfy the following inequality constraints:
[0014] in The K-1th output node of the model is at the q Dimensional network residuals, i.e. The q-th column, where This constraint serves as the supervision mechanism for the semi-supervised randomized network, ensuring that the weights and biases of newly added hidden nodes enable the semi-supervised randomized network to converge quickly. Nodes that satisfy the above inequality constraint are added to the node candidate pool.
[0015] Furthermore, S6 specifically includes: defining a set of variables Calculate the nodes in the candidate pool The neural node with the maximum value is selected as the optimal node.
[0016] Furthermore, S7 specifically includes: initialization settings ; S8. Specifically, this includes calculating the output weights as follows:
[0017] Unlabeled indicator matrix Perform the following operation on each column:
[0018] in, It is the first unlabeled matrix i Line number j List, , for The j List, If the sum of each row is 1, then The solution can be found in the following way:
[0019] because ,so Since it is a convex function, it can be solved using Newton's method. The statement is as follows:
[0020]
[0021] C is the change, C and ft Control the loop of iterative optimization; Repeat S8 until the constraints are met.
[0022] Compared to existing technologies, the principles and advantages of this solution are as follows: 1. This invention provides a modeling method and system for semi-supervised randomized network, which can effectively utilize unlabeled sample information, adaptively adjust weights and biases according to training data, eliminate backpropagation process, avoid the problems of slow training speed and getting trapped in local optima caused by gradient descent, have low hardware requirements, and achieve accurate identification of training samples.
[0023] 2. This invention employs a manifold regularization strategy to obtain the geometric similarity relationship between unlabeled and labeled samples, and uses the structural information of unlabeled samples to assist in the modeling of semi-supervised randomized network.
[0024] 3. This invention innovatively introduces a joint optimization learning strategy into a semi-supervised randomized network, enabling iterative optimization of the output weight matrix and the indicator matrix of unlabeled samples to enhance its ability to mine local information and improve the recognition accuracy of the algorithm.
[0025] 4. This invention designs a novel inequality constraint mechanism to adaptively allocate the optimal weights and biases of hidden nodes. It requires minimal manual parameter adjustment and theoretically guarantees the global approximation property of the model.
[0026] 5. Three industrial fault diagnosis examples demonstrate that the proposed semi-supervised randomized configuration network outperforms seven other outstanding classification algorithms, proving the effectiveness of the method and showing that it has good fault identification performance in industry. Attached Figure Description
[0027] Figure 1 This is a flowchart illustrating an embodiment of a fault diagnosis method based on a semi-supervised random configuration network according to the present invention. Figure 2 This is a schematic diagram of an embodiment of a fault diagnosis method based on a semi-supervised random configuration network according to the present invention. Figure 3 This is a schematic diagram of the TEP experimental simulation process of an embodiment of the fault diagnosis method based on a semi-supervised random configuration network according to the present invention. Figure 4 This is a schematic diagram of the confusion matrix of a TEP (Trouble Diagnosis Method) based on a semi-supervised random configuration network according to the present invention. Figure 5 This is a schematic diagram of the confusion matrix for JOSSCN diagnosis, an embodiment of the fault diagnosis method based on a semi-supervised random configuration network according to the present invention. Figure 6 This is a flowchart illustrating an embodiment of the fault diagnosis method CWRU based on a semi-supervised random configuration network according to the present invention. Figure 7 This is a time-frequency diagram of CWT under different states in an embodiment of a fault diagnosis method based on a semi-supervised random configuration network according to the present invention. Figure 8 This is a schematic diagram of the confusion matrix of CWRU, an embodiment of the fault diagnosis method based on a semi-supervised random configuration network according to the present invention. Detailed Implementation
[0028] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 and attached Figure 2 As shown: A fault diagnosis method based on semi-supervised randomized network configuration includes the following steps: S1. Vibration signal acquisition is performed using specialized equipment for key components of industrial equipment; S2. Data preprocessing: Analyze the collected signals and extract features; standardize the data to eliminate the influence of dimensions, thereby improving the model training effect; S3. Initialize the semi-supervised randomized network: Split the network into unlabeled samples, labeled samples, and a test set; set the Laplacian matrix parameters and construct the Laplacian matrix; determine the optimal number of hidden layer nodes. Critical tolerance error Number of candidate nodes Set the range of random parameters and ; S4. Randomly configure the weights and biases of the nodes and construct candidate neural nodes for the hidden layer; S5. Select candidate nodes using inequality constraints; S6. Determine the optimal neural node and add it to the network model; S7. Initialize the unlabeled sample indicator matrix; S8. Calculate the output weight matrix and use a joint optimization method to iteratively optimize the index matrix and output weight matrix of the unlabeled samples; S3 specifically includes: given a set of preprocessed data, inputting unlabeled samples, labeled samples, and a test set; initializing a semi-supervised randomized network. Input dataset ,in For sample features, For sample labels, ,in For labeled sample features, To indicate the number of tags, This represents the number of unlabeled samples. , For X L The tag, Each element in is equal to Therefore The training process is unknown.l + u = n Initialize the Laplace matrix parameters and determine the optimal number of hidden layer nodes. Critical tolerance error Number of candidate nodes ,set up The selection range is , r The selection range is Regularization coefficient and The selection range is as follows: The optimization objective of the semi-supervised randomized network is as follows:
[0029]
[0030]
[0031] in h For activation function, This is the output of the Kth hidden node. For Laplace matrix, The calculation method is as follows:
[0032] in yes x of k A collection of fields The average edge length of the local adjacency graph .
[0033] S4 specifically includes: from In the process, output weights are randomly selected. and bias From a sequence of nonnegative real numbers Select r Substitute into the activation function get Group candidate neural nodes .
[0034] S5 specifically includes: The semi-supervised randomized network is a single-hidden-layer network, and the network model is built by adding nodes one by one. Assuming that a network model containing K-1 hidden-layer nodes has already been built, the newly added Kth neuron node in the semi-supervised randomized network needs to satisfy the following inequality constraints:
[0035] in The K-1th output node of the model is at the q Dimensional network residuals, i.e. The q-th column, where This constraint serves as the supervision mechanism for the semi-supervised randomized network, ensuring that the weights and biases of newly added hidden nodes enable the semi-supervised randomized network to converge quickly. Nodes that satisfy the above inequality constraint are added to the node candidate pool.
[0036] S6 specifically includes: defining a set of variables: Calculate the nodes in the candidate pool The neural node with the maximum value is selected as the optimal node.
[0037] The larger the value of , the faster the convergence speed. The mathematical proof of the reason for this convergence is as follows:
[0038] In the formula for The function, The newly added node is determined by random parameters. Taking the derivative of the above equation and setting it to 0, we can derive the following:
[0039] Based on the above formula, the error attenuation factor The derivation is as follows:
[0040] If the existing hidden nodes remain unchanged, then It is a fixed value. To make the attenuation factor... To maximize, one should choose an appropriate... Make minimize.
[0041] When the number of nodes in the model is large enough, the constraints will guarantee that the model converges strictly, as proven below:
[0042] in Let K-1 be the number of nodes and represent the model error. season Then the above formula can be derived as follows:
[0043] make Then we can obtain the following formula:
[0044] If the added node guarantees the following constraints, then the model will inevitably converge gradually, and the obtained... The larger the value, the faster the convergence speed.
[0045] S7 specifically includes: initialization settings ; S8. Specifically, this includes calculating the output weights as follows:
[0046] Unlabeled indicator matrix Perform the following operation on each column:
[0047] in, It is the first unlabeled matrix i Line number j List, , for The j List, If the sum of each row is 1, then The solution can be found in the following way:
[0048] because ,so Since it is a convex function, it can be solved using Newton's method. The statement is as follows:
[0049]
[0050] C is the change, C and ft Control the loop of iterative optimization; Repeat step S8 until the constraints are satisfied. That is what I wanted; Compared to existing technologies, the principles and advantages of this solution are as follows: 1. This invention provides a modeling method and system for semi-supervised randomized network, which can effectively utilize unlabeled sample information, adaptively adjust weights and biases according to training data, eliminate backpropagation process, avoid the problems of slow training speed and getting trapped in local optima caused by gradient descent, have low hardware requirements, and achieve accurate identification of training samples.
[0051] 2. This invention employs a manifold regularization strategy to obtain the geometric similarity relationship between unlabeled and labeled samples, and uses the structural information of unlabeled samples to assist in the modeling of semi-supervised randomized network.
[0052] 3. This invention innovatively introduces a joint optimization learning strategy into a semi-supervised randomized network, enabling iterative optimization of the output weight matrix and the indicator matrix of unlabeled samples to enhance its ability to mine local information and improve the recognition accuracy of the algorithm.
[0053] 4. This invention designs a novel inequality constraint mechanism to adaptively allocate the optimal weights and biases of hidden nodes. It requires minimal manual parameter adjustment and theoretically guarantees the global approximation property of the model.
[0054] 5. Three industrial fault diagnosis examples demonstrate that the proposed semi-supervised randomized configuration network outperforms seven other outstanding classification algorithms, proving the effectiveness of the method and showing that it has good fault identification performance in industry.
[0055] This experimental implementation case uses publicly available laboratory datasets as the source of the original data. Specifically, the Tennessee Eastman Process (TEP) dataset, the Air Compressor Working Process (ACWP) dataset, and the Case Western Reserve University (CWRU) dataset are used as experimental cases. The data labels are required to be complete and the amount of data for each state should be balanced.
[0056] We performed a performance comparison analysis of our proposed algorithm (JOSSCN) with SVM, RVFL, SCN, SS-RVFL, JOSRVFL, and LPSCN on the three datasets mentioned above. The regularization parameters were all within the range of... The selection of neural nodes for all models is within the range of {25:25:1000}. Accuracy, Precision, Recall, and F1 Score are used as performance evaluation metrics. The results of all algorithms are run 20 times and the average value is taken.
[0057] Case 1: Tennessee Eastman's processing procedures, such as Figure 3 As shown, the process mainly consists of five components: reactor, product condenser, recovery compressor, product separator, and gas-liquid separator. The TEP dataset is a classic dataset for control and fault diagnosis of chemical processes, exhibiting strong coupling and nonlinear characteristics. Ten equipment states were selected from the TEP data, and the data samples possess 52 features.
[0058] S1. Sensor signal acquisition: 11 operational variables and 41 process variables were acquired through Matlab simulation. S2. Data Preprocessing: The standard deviation standardization method is used to preprocess the samples to eliminate the influence between variables and achieve dimensionless data. S3. Initialize the semi-supervised randomized network with 800 unlabeled samples, 4000 labeled samples, and 3200 test samples. Set the Laplacian matrix parameters and construct the Laplacian matrix. Set the optimal number of hidden layer nodes. Critical tolerance error Number of candidate nodes Set the range of random parameters and ; S4. Randomly configure the weights and biases of the nodes and construct candidate neural nodes for the hidden layer; S5. Select candidate nodes using inequality constraints; S6. Determine the optimal neural node and add it to the network model; S7. Initialize the unlabeled sample indicator matrix; S8. Calculate the output weight matrix and use a joint optimization method to iteratively optimize the index matrix and output weight matrix of the unlabeled samples; S9. Input the test set and output the diagnostic results.
[0059] As shown in Table 1, JOSSCN exhibits the best performance among all methods, with an accuracy of 85.47%, precision of 86.79%, recall of 85.48%, and F1-score of 85.56%. The experimental results outperform SVM, RVFL, SCN, and GSCN, validating the effectiveness of the proposed method. The confusion matrix for JOSSCN diagnosis is shown below. Figure 4 As shown.
[0060] Table 1
[0061] Case 2: An air compressor is a device that converts electrical energy into gas pressure energy. It is widely used in manufacturing, painting, pneumatic control, mining, and construction industries. The ACWP dataset consists of 1800 samples, including 8 states: one healthy state and 7 fault states, namely: leaking intake valve fault, leaking exhaust valve fault, check valve fault, piston ring fault, flywheel fault, drive belt fault, and bearing fault.
[0062] S1. Sensor signal acquisition: The unidirectional microphone is 1.5cm away from the air compressor. The entire process is performed using a single NI 9234 data acquisition (DAQ) hardware unit, an NI 9172 USB interface, and a LabVIEW-based data acquisition interface. S2. Data Preprocessing: Wavelet Packet Transform feature extraction was used to extract 254 data features from the signal. Then, the standard deviation standardization method was used to preprocess the sample to eliminate the influence between variables and achieve dimensionless data. S3. Initialize the semi-supervised randomized network with 902 unlabeled samples, 32 labeled samples, and 866 test samples. Set the Laplacian matrix parameters and construct the Laplacian matrix. Set the optimal number of hidden layer nodes. Critical tolerance error Number of candidate nodes Set the range of random parameters and ; S4. Randomly configure the weights and biases of the nodes and construct candidate neural nodes for the hidden layer; S5. Select candidate nodes using inequality constraints; S6. Determine the optimal neural node and add it to the network model; S7. Initialize the unlabeled sample indicator matrix; S8. Calculate the output weight matrix and use a joint optimization method to iteratively optimize the index matrix and output weight matrix of the unlabeled samples; S9. Input the test set and output the diagnostic results.
[0063] Table 2 shows that JOSSCN exhibits the best performance among all methods, with an accuracy of 91.04%, precision of 91.79%, recall of 91.04%, and F1-score of 91.06%. The experimental results outperform SVM, RVFL, SCN, and GSCN, validating the effectiveness of the proposed method. The confusion matrix for JOSSCN diagnosis is shown below. Figure 5 As shown.
[0064] Table 2
[0065] Case 3 The CWRU dataset covers 10 vibration scenarios for drive-end bearings under normal vibration conditions and three fault types: rolling element failure, inner raceway failure, and outer ring failure (fault diameters of 0.007 inches, 0.014 inches, and 0.021 inches, respectively). The process in this case is as follows: Figure 6 As shown.
[0066] 1. The accelerometer collects the vibration signal of the drive end bearing when the motor load is 2 horsepower at a sampling rate of 12,000 Hz; 2. Data Preprocessing: The original signal is segmented, and continuous wavelet transform (CWT) is used to generate time-frequency graphs of the bearing fault vibration signal. CWT time-frequency images under different states are shown below. Figure 7 As shown, the gray-level difference statistical features of the image are then calculated, and nine data features are extracted as input to the algorithm. The standard deviation standardization method is then used to preprocess the samples to eliminate the influence between variables and realize the dimensionless data. 3. Initialize the semi-supervised randomized network with 600 unlabeled samples, 20 labeled samples, and 580 test samples. Set the Laplacian matrix parameters and construct the Laplacian matrix. Determine the optimal number of hidden layer nodes. Critical tolerance error Number of candidate nodes Set the range of random parameters and ; 4. Randomly configure the weights and biases of nodes, and construct candidate neural nodes for the hidden layer; 5. Select candidate nodes using inequality constraints; 6. Determine the optimal neural node and add it to the network model; 7. Initialize the unlabeled sample indicator matrix; 8. Calculate the output weight matrix, and use a joint optimization method to iteratively optimize the index matrix and output weight matrix of the unlabeled samples; 9. Input the test set and output the diagnostic results.
[0067] Table 3 shows that JOSSCN exhibits the best performance among all methods, with an accuracy of 95.44%, precision of 95.76%, recall of 95.44%, and F1-score of 95.37%. The experimental results outperform SVM, RVFL, SCN, and GSCN, validating the effectiveness of the proposed method. The diagnostic results of JOSSCN are as follows: Figure 8 As shown.
[0068] Table 3
[0069] The above is merely an implementation of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A fault diagnosis method based on a semi-supervised randomized network, characterized in that, Fast and efficient modeling is achieved through a semi-supervised stochastic network. This method first preprocesses the collected data, then initializes the network parameters, and then uses a supervision mechanism to construct a pool of candidate neural nodes. During this process, the best candidate node is selected as the hidden layer node. Next, the output weight matrix is calculated, and the output weight matrix is jointly iteratively optimized with the unlabeled sample indicator matrix. Finally, the model performance is evaluated based on the fault identification accuracy. The method includes the following steps: S1. Vibration signal acquisition is performed using specialized equipment for key components of industrial equipment; S2. Data preprocessing: Analyze the collected signals and extract features; standardize the data to eliminate the influence of dimensions, thereby improving the model training effect; S3. Initialize the semi-supervised randomized network: Split the network into unlabeled samples, labeled samples, and a test set; set the Laplacian matrix parameters and construct the Laplacian matrix; determine the optimal number of hidden layer nodes. Critical tolerance error Number of candidate nodes Set the range of random parameters and ; S4. Randomly configure the weights and biases of the nodes and construct candidate neural nodes for the hidden layer; S5. Select candidate nodes using inequality constraints; S6. Determine the optimal neural node and add it to the network model; S7. Initialize the unlabeled sample indicator matrix; S8. Calculate the output weight matrix and use a joint optimization method to iteratively optimize the index matrix and output weight matrix of the unlabeled samples; S9. Perform fault diagnosis on the test sample; S3 specifically includes: given a set of preprocessed data, inputting unlabeled samples, labeled samples, and a test set; initializing a semi-supervised randomized network. Input dataset ,in For sample features, For sample labels, ,in For labeled sample features, To indicate the number of tags, This represents the number of unlabeled samples. , For X L The tag, Each element in is equal to Therefore The training process is unknown. l + u = n Initialize the Laplace matrix parameters and determine the optimal number of hidden layer nodes. Critical tolerance error Number of candidate nodes ,set up The selection range is , r The selection range is Regularization coefficient and The selection range is as follows: The optimization objective of the semi-supervised randomized network is as follows: in h For activation function, This is the output of the Kth hidden node. For Laplace matrix, The calculation method is as follows: in yes x of k A collection of fields The average edge length of the local adjacency graph 。 2. The fault diagnosis method based on a semi-supervised randomized network according to claim 1, characterized in that: S4 specifically includes: from In the process, output weights are randomly selected. and bias From the sequence of nonnegative real numbers Select r Substitute into the activation function get Group candidate neural nodes .
3. The fault diagnosis method based on a semi-supervised randomized network according to claim 1, characterized in that: S5 specifically includes: The semi-supervised randomized network is a single-hidden-layer network, and the network model is built by adding nodes one by one. Assuming that a network model containing K-1 hidden-layer nodes has already been built, the newly added Kth neuron node in the semi-supervised randomized network needs to satisfy the following inequality constraints: in The K-1th output node of the model is at the q Dimensional network residuals, i.e. The qth column, where This constraint serves as the supervision mechanism for the semi-supervised randomized network, ensuring that the weights and biases of newly added hidden nodes enable the semi-supervised randomized network to converge quickly. Nodes that satisfy the above inequality constraint are added to the node candidate pool.
4. The fault diagnosis method based on a semi-supervised randomized network according to claim 1, characterized in that: S6 specifically includes: defining a set of variables: Calculate the nodes in the candidate pool The neural node with the highest value is selected as the optimal node.
5. The fault diagnosis method based on a semi-supervised randomized network according to claim 1, characterized in that: S7 specifically includes: initialization settings ; S8. Specifically, this includes calculating the output weights as follows: Unlabeled indicator matrix Perform the following operation on each column: in, It is the first unlabeled matrix i Line number j List, , for The j List, If the sum of each row is 1, then The solution can be found in the following way: because ,so Since it is a convex function, it can be solved using Newton's method. The statement is as follows: C is the change, C and ft Control the loop of iterative optimization; Repeat S8 until the constraints are met.
6. A fault diagnosis system for a semi-supervised randomized network using the fault diagnosis method as described in any one of claims 1 to 5, characterized in that, include: The signal acquisition module is used for data acquisition. The data preprocessing module is used for signal analysis, feature extraction, and standardization of data. The initialization module is used to initialize the parameters of a semi-supervised, randomly configured network. The random configuration module is used to configure the input weights and biases of nodes; The hidden layer node selection module is used to filter nodes and select the best node; Output weight joint optimization module; used to calculate network output weights and perform joint iterative optimization with the unlabeled sample indicator matrix.