Fault diagnosis method based on feed-forward network and sample combination strategy
By proposing a fault diagnosis method based on feedforward networks and sample combination strategies, this paper addresses the problems of coarse application and reliance on complex annotations in fault diagnosis using FF networks. It achieves efficient and accurate fault diagnosis, reduces sample annotation costs, and improves the reliability and engineering applicability of diagnostic results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-10
- Publication Date
- 2026-05-15
AI Technical Summary
Existing FF networks are crudely applied in fault diagnosis, rely on complex annotations, have insufficient diagnostic accuracy, and their advantages in parameter interpretability have not been explored, making it difficult to meet actual industrial needs.
By adopting a feedforward network and sample combination strategy, training samples are constructed through data as labels. Combined with frequency domain data preprocessing and goodness function optimization objective, a feedforward network containing two fully connected layers is constructed to achieve efficient fault diagnosis without complex annotation.
It achieves efficient and accurate fault diagnosis, reduces sample annotation costs, improves the reliability and engineering applicability of diagnostic results, and is suitable for industrial scenarios where fault samples are scarce.
Smart Images
Figure CN122045947A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mechanical equipment fault diagnosis technology, specifically relating to a fault diagnosis method based on feedforward network and sample combination strategy. It is applicable to the accurate diagnosis of known fault categories of rotating machinery such as bearings and gears, and is especially suitable for industrial scenarios where fault sample acquisition is limited and data annotation costs are high. Background Technology
[0002] Rotating machinery is a core piece of equipment in modern industrial production, and the operating status of key components such as bearings and gears directly affects the continuity and safety of production. These components operate under complex conditions of high frequency, load fluctuations, and strong noise interference for extended periods, making them prone to wear, cracks, and broken teeth. If these issues are not diagnosed promptly and accurately, they can lead to equipment downtime, production interruptions, and even safety accidents, resulting in significant economic losses.
[0003] Among existing methods for diagnosing rotating machinery faults, deep learning-based diagnostic techniques are widely used due to their automatic feature extraction capabilities. Backpropagation (BP) networks are one of the mainstream models, but they suffer from problems such as complex parameter updates, susceptibility to local optima, and extremely high dependence on sample labeling. In real-world industrial scenarios, fault samples are scarce, and labeling requires specialized personnel, limiting the diagnostic accuracy and generalization ability of BP networks.
[0004] Feedforward (FF) networks, with their advantages of simple structure, efficient parameter updates, and fast convergence, have potential applications in pattern recognition. Unlike the "black box" nature of traditional backpropagation (BP) networks, they possess inherent parameter interpretability. However, current technologies have very limited applications of FF networks in fault diagnosis. Their parameter interpretability advantage has not been fully utilized, and traditional sample annotation methods are often directly adopted without considering the frequency domain characteristics of fault data to design targeted sample construction strategies. This hinders the realization of their core advantages, and the diagnostic accuracy falls short of actual industrial needs. Therefore, there is an urgent need for a method that adapts to the characteristics of FF networks, eliminates the need for complex annotations, and relies on sample combinations to achieve efficient fault diagnosis, fully leveraging the structural advantages and interpretability value of FF networks to improve the accuracy and engineering practicality of fault diagnosis. Summary of the Invention
[0005] To address the technical problems of existing feedforward networks in fault diagnosis, such as their extensive application, reliance on complex annotations, insufficient diagnostic accuracy, and lack of exploitation of parameter interpretability advantages, this invention provides a fault diagnosis method based on feedforward networks and a sample combination strategy. By constructing training samples through a data-as-label sample combination strategy, and adapting to the structural characteristics and parameter interpretability advantages of feedforward networks, this method achieves accurate diagnosis of known fault categories, reduces sample annotation costs, and enhances the reliability of diagnostic results by relying on interpretability, thereby strengthening the engineering applicability of the method.
[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: A fault diagnosis method based on feedforward networks and sample combination strategies, the specific steps of which are as follows:
[0007] Step 1: Use sensors to collect bearing operating data, obtaining data on normal operation and various types of faults;
[0008] Step 2: Perform Fourier transform on the data and then perform min-max normalization;
[0009] Step 3: Combine the data of the same type obtained in Step 2 into pairs to form positive sample pairs, and combine the data of different types into negative sample pairs;
[0010] Step 4: Construct a feedforward network containing two fully connected layers. Input positive and negative sample pairs into the first layer network in sequence, calculate the goodness of performance using the output of the first layer, and update the parameters of the first layer network based on the goodness of performance.
[0011] Step 5: Use the output of the first layer as the input of the second layer to obtain the output of the second layer and calculate the goodness of the second layer. Update the parameters of the second layer according to the goodness of the second layer.
[0012] Step 6: Combine the test data of the bearing to be tested with the existing data to form sample pairs, and input them into the trained feedforward network. Use the output of the second layer network to calculate the goodness of fit, and determine the type of bearing fault based on the goodness of fit.
[0013] In step one, the bearing operation data is collected using sensors to obtain dataset X. The specific process includes the following steps:
[0014] Step S11: Install the vibration signal sensor onto the bearing, ensuring that the sensor is correctly connected and can work properly;
[0015] Step S12: Collect the operating data of bearings with C types of faults, and record them as follows: .
[0016] In step two, the data undergoes Fourier transform and min-max normalization. The specific process includes the following steps:
[0017] Step S21: Divide the dataset X into multiple data samples according to N data points as a group. For each sample, the Discrete Fourier Transform is used to eliminate noise interference in the time-domain signal. The specific formula is as follows:
[0018] ;
[0019] Where N is the number of sampling points, k=0,1,…N-1, The data is in complex form in the frequency domain, containing amplitude and phase information, where j is the imaginary unit ( ), The rotation factor is used to decompose the discrete-time signal into sine / cosine components of different frequencies, perform the discrete Fourier transform described above, and calculate the frequency domain amplitude. And retain the positive frequency components. After removing the negative frequency components, one-dimensional frequency domain data y is obtained;
[0020] Step S22: Standardize the frequency domain data y using the min-max normalization method to map the data amplitude to the [0,1] interval. The specific formula is as follows:
[0021] .
[0022] In step S3, pairs of data of the same type are combined to form positive sample pairs. Different types of data combinations constitute negative sample pairs The specific process includes the following steps:
[0023] Step S31: Employing a data-as-label strategy, no manual labeling of fault categories is required. Positive and negative sample pairs are constructed through data overlay. Category association is achieved based on the consistency of the frequency domain characteristics of the fault vibration signals. Vibration signals in the same state (normal or specific fault) exhibit significant consistency in their frequency domain characteristic distribution after discrete Fourier transform, while signals in different states show significant differences in their frequency domain characteristics. This characteristic provides a foundation for automatic sample classification. Specifically, two... Let them be vectors respectively. sum vector The input sample is obtained by adding corresponding elements of the two vectors. (in If these two If the input samples belong to the same fault category, the constructed input samples are positive sample pairs, denoted as . Otherwise, it is a negative sample and is recorded as .
[0024] Step S32, Sample set division: Divide the constructed positive and negative sample pairs into training and test sets in a 1:9 ratio. The training set is used for training the parameters of the feedforward network model, and the test set is used to verify the diagnostic accuracy of the model. This ensures that the number of positive and negative samples corresponding to various faults in the sample set is balanced, and avoids model training bias.
[0025] In step four, a feedforward network containing two fully connected layers is constructed. Positive and negative sample pairs are sequentially input into the first layer network. The output of the first layer network is used to calculate the goodness of performance, and the parameters of the first layer network are updated based on the goodness of performance. The specific process includes the following steps:
[0026] Step S41: The feedforward network model consists of two fully connected layers, the difference being the number of units. The initial weight matrix of each layer adopts a random normal distribution, and the activation function is the Sigmoid function, which is used to achieve non-linear mapping of features.
[0027] Step S42: First, train the first layer of the network. The output vector of the first layer, denoted as L, is obtained by inputting sample pairs into the network. The goodness-of-performance (OG) of L is calculated as the optimization objective for model training. OG is defined as:
[0028] ;
[0029] in, For the goodness function, It is the Sigmoid activation function. This is the output value of the first layer;
[0030] Step S43: During training, maximize the goodness value of positive samples and minimize the goodness value of negative samples. Use a stochastic gradient descent optimizer to update the network weight parameters. The weight update formula is:
[0031] ;
[0032] in, For learning rate, The weights of the first layer samples, Here, G represents the updated weights of the first layer of the network, and G is the goodness function. For the goodness function pair The derivative of , where sign is the sign coefficient, taking 1 for positive sample pairs and -1 for negative sample pairs.
[0033] In step five, the output L of the first layer is used as the input of the second layer to obtain the output P of the second layer, and the goodness is calculated. The parameters of the second layer are then updated based on the goodness. The specific process includes the following steps:
[0034] Step S51: Input L into the second layer network to obtain the output vector of the second layer, denoted as P. Calculate the goodness of performance based on P as the optimization objective for training the second layer. The goodness of performance is defined as:
[0035] ;
[0036] in, For the goodness function, It is the Sigmoid activation function. This is the output value of the second layer;
[0037] Step S52: During training, maximize the goodness value of positive samples and minimize the goodness value of negative samples. Use a stochastic gradient descent optimizer to update the network weight parameters. The weight update formula is:
[0038] ;
[0039] in, For learning rate, The weights of the first layer samples, Here, G represents the updated weights of the second-layer network, and G is the goodness function. For the goodness function pair The derivative of , sign is the coefficient with sign, 1 for positive sample pairs and -1 for negative sample pairs. Repeat steps S42-S52, training two layers of network in turn as one round, setting the number of training rounds to 120 rounds, and the batch size to 8-16. After each round of training, calculate the accuracy of the validation set. When the accuracy has been trained for 3 consecutive rounds and the fluctuation does not exceed 1%, stop training and save the optimal model parameters.
[0040] In step six, the bearing inspection data to be tested is sequentially paired with existing data to form sample pairs, which are then input into the trained feedforward network. The output of the second layer network is used to calculate the goodness of fit, and the bearing fault type is determined based on the goodness of fit. The specific process includes the following steps:
[0041] Step S61: Construct input sample pairs: Perform discrete Fourier transform and normalization on the samples to be detected, combine them with the data of each category of historical data, input each sample pair into the discrete network, and the model outputs the goodness value of each sample pair.
[0042] Step S62, Fault Category Determination: Identify the sample pair with the highest excellence value. The corresponding category of the historical data contained in this sample pair is the fault diagnosis result.
[0043] Compared with the prior art, the specific beneficial effects of this invention are reflected in:
[0044] I. High diagnostic efficiency and strong parameter interpretability, adapted to feedforward network characteristics: This invention fully utilizes the advantages of feedforward networks, such as simple structure and no backpropagation link, combined with a simplified sample construction strategy, resulting in efficient parameter updates and rapid model training to meet the real-time diagnostic needs of industrial sites. At the same time, feedforward networks have clear parameter interpretability, and the network weight matrix can directly map the effective fault feature frequencies in the frequency domain data, clearly tracing the source of diagnostic results and identifying which frequency components play a key role in fault identification. This not only facilitates fault mechanism analysis but also improves the credibility of diagnostic results, solving the drawbacks of black-box diagnosis in traditional deep learning models.
[0045] Second, it reduces sample dependence and eliminates the need for complex labeling: The present invention adopts a data-as-label sample combination strategy, which eliminates the need for professional personnel to manually label fault categories. Training samples can be constructed simply by matching data states, which greatly reduces the cost of sample labeling and is suitable for industrial scenarios where fault samples are scarce and labeling is difficult.
[0046] III. High diagnostic accuracy and strong generalization ability: This invention eliminates noise interference through frequency domain data preprocessing and optimizes the target by combining the goodness function, enabling the feedforward network to accurately capture the characteristic differences of different fault categories. Experimental verification shows that the diagnostic accuracy of known fault categories can reach more than 95%, and it still maintains stable performance under different load conditions.
[0047] IV. Strong Engineering Practicality: The network structure of this invention is simple and the training process is convenient. It does not require complex hardware support and can be directly integrated into the vibration monitoring system in the industrial field. It is suitable for the fault diagnosis needs of various rotating machinery such as bearings and gears and is easy to promote and apply. Attached Figure Description
[0048] Figure 1 This is a schematic diagram comparing the feedforward network structure and the backpropagation network structure used in this invention.
[0049] Figure 2 This is a schematic diagram illustrating the principle of data, i.e., label sample combination in this invention.
[0050] Figure 3 This is the confusion matrix of the feedforward network diagnostic results in Example 1.
[0051] Figure 4 This is a visualization comparison of the parameters of the first layer of the feedforward network in Example 1.
[0052] Figure 5 This is the confusion matrix of the feedforward network diagnostic results in Example 2.
[0053] Figure 6 This is a visualization comparison of the parameters of the first layer of the feedforward network in Example 2. Detailed Implementation
[0054] To make the technical problem to be solved, the technical solution, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and two bearing failure experimental cases. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0055] This embodiment uses the CWRU (Case Western Reserve University) bearing dataset as the experimental object to verify the effectiveness of the proposed method. The CWRU dataset contains 10 categories: normal bearing condition (C0), inner race fault (C1), outer race fault (C2), and rolling element fault (C3). Each fault category includes three different severity levels. The experimental equipment is a motor bearing test bench, using an accelerometer with a sampling frequency of 12kHz. The composite fault dataset originates from the 2009 Predictive and Health Management Data Challenge and includes 8 categories: normal condition (C0) and composite faults (C1-C7). The experimental equipment is a gearbox, using an accelerometer with a sampling frequency of 66.67kHz.
[0056] Example 1: Fault diagnosis experiment based on CWRU bearing dataset:
[0057] Data acquisition and preprocessing: Bearing vibration signals under ten conditions were acquired, with 50 sets of samples for each condition, and each set of samples containing 2048 data points; Discrete Fourier transform was performed on each set of vibration signals to convert them into one-dimensional frequency domain data, and the data was standardized to the [0,1] interval using the minimum-maximum normalization method to eliminate amplitude differences.
[0058] Sample pair construction: strictly follows Figure 2 The sample combination principle shown employs a "data as label" strategy to construct positive and negative sample pairs, which are divided into training and test sets in a 1:9 ratio to ensure a balanced sample class and provide a reliable data foundation for model training.
[0059] FF network model construction: The FF network consists of two fully connected layers with 100 and 50 neurons respectively; the learning rate is 0.001, the activation function is Sigmoid, the optimizer is SGD, the batch size is 10, and the number of training epochs is 120.
[0060] Model Training and Diagnosis: The training set is input into the feedforward network model for training. The training process is as follows: Figure 3 As shown, as training progresses, the recognition accuracy of various faults gradually reaches a higher level. The average accuracy stabilizes after the 100th round of training, and the final training set accuracy reaches 100%. Figure 4 As shown, the interpretability advantage of the feedforward network parameters is fully demonstrated. Analysis of the weight matrix distribution reveals that the weight values corresponding to the bearing fault characteristic frequencies are significantly higher than other frequency components, which perfectly matches the bearing fault mechanism, achieving closed-loop verification of "diagnosis results - network parameters - fault mechanism". Inputting the test set into the trained model, the diagnostic results show an average diagnostic accuracy of 100% for the four states, verifying the accuracy of this method and highlighting the auxiliary value of the feedforward network's interpretability in fault diagnosis.
[0061] Example 2: Fault diagnosis experiment based on composite fault bearing dataset:
[0062] Data acquisition and preprocessing: Bearing vibration signals under eight conditions were acquired, with 50 sets of samples for each condition and 4000 data points in each set. Discrete Fourier transform was performed on each set of vibration signals to convert them into one-dimensional frequency domain data. The minimum-maximum normalization method was used to standardize the data to the [0,1] interval to eliminate amplitude differences.
[0063] Sample pair construction: strictly follows Figure 2 The sample combination principle shown employs a "data as label" strategy to construct positive and negative sample pairs, which are divided into training and test sets in a 1:9 ratio to ensure a balanced sample class and provide a reliable data foundation for model training.
[0064] Feedforward network model construction: The feedforward network consists of two fully connected layers with 100 and 50 neurons respectively; the learning rate is 0.001, the activation function is Sigmoid, the optimizer is SGD, the batch size is 10, and the number of training epochs is 120.
[0065] Model Training and Diagnosis: The training set is input into the feedforward network model for training. The training results are as follows: Figure 5 As shown, the final test set accuracy reached 100%. Due to the complex fault generation mechanism, it is difficult to calculate the fault characteristic frequencies. However, all types of data contain some common frequencies, such as rotational frequency (8Hz) and gear meshing frequency (320Hz, 640Hz), which play a negative role in identification. Figure 6 The diagram illustrates that the feedforward network marks the gear meshing frequency in black to indicate that it is a negative frequency. Due to the limited number of input samples (4000), the frequency resolution cannot reach 8Hz, and the diagram does not provide any frequency conversion information. However, the second harmonic frequency with a small amplitude (16Hz) is also defined as a negative frequency, which further demonstrates the good interpretability of the feedforward network.
[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the scope of the present invention.
Claims
1. A fault diagnosis method based on feedforward networks and sample combination strategies, characterized in that, The specific steps are as follows: Step 1: Use sensors to collect bearing operating data, obtaining data on normal operation and various types of faults; Step 2: Perform Fourier transform on the data and then perform min-max normalization; Step 3: Combine the data of the same type obtained in Step 2 into pairs to form positive sample pairs, and combine the data of different types into negative sample pairs; Step 4: Construct a feedforward network containing two fully connected layers. Input positive and negative sample pairs into the first layer network in sequence, calculate the goodness of performance using the output of the first layer, and update the parameters of the first layer network based on the goodness of performance. Step 5: Use the output of the first layer as the input of the second layer to obtain the output of the second layer and calculate the goodness of the second layer. Update the parameters of the second layer according to the goodness of the second layer. Step 6: Combine the test data of the bearing to be tested with the existing data to form sample pairs, and input them into the trained feedforward network. Use the output of the second layer network to calculate the goodness of the bearing and determine the type of bearing failure based on the goodness of ...
2. The fault diagnosis method based on feedforward network and sample combination strategy according to claim 1, characterized in that, In step one, the bearing operation data is collected using sensors to obtain dataset X. The specific process includes the following steps: Step S11: Install the vibration signal sensor onto the bearing, ensuring that the sensor is correctly connected and can work properly; Step S12: Collect the operating data of bearings with C types of faults, and record them as follows: .
3. The fault diagnosis method based on feedforward network and sample combination strategy according to claim 2, characterized in that, In step two, the data undergoes Fourier transform and min-max normalization. The specific process includes the following steps: Step S21: Divide the dataset X into multiple data samples according to N data points as a group. For each sample, a discrete Fourier transform is used to eliminate noise interference in the time-domain signal; Step S22: Use the min-max normalization method to standardize the frequency domain data so that the data amplitude is mapped to the [0,1] interval.
4. The fault diagnosis method based on feedforward network and sample combination strategy according to claim 3, characterized in that, In step S3, pairs of data of the same type are combined to form positive sample pairs. Different types of data combinations constitute negative sample pairs The specific process includes the following steps: Step S31: Employing a data-as-label strategy, positive and negative sample pairs are constructed through data overlay. Category association is achieved based on the consistency of the frequency domain characteristics of the fault vibration signals. Two samples are selected... Let them be vectors respectively. sum vector The input sample is obtained by adding corresponding elements of the two vectors. ,in, If two If the input samples belong to the same fault category, then the constructed input samples are positive sample pairs, denoted as . If two If the samples do not belong to the same fault category, they are recorded as negative samples. ; Step S32, Sample set division: Divide the constructed positive sample pairs and negative sample pairs into training set and test set in a 1:9 ratio. The training set is used for training the parameters of the feedforward network model, and the test set is used to verify the diagnostic accuracy of the model.
5. The fault diagnosis method based on feedforward network and sample combination strategy according to claim 4, characterized in that, In step four, a feedforward network containing two fully connected layers is constructed. Positive and negative sample pairs are sequentially input into the first layer network. The output of the first layer network is used to calculate the goodness of performance, and the parameters of the first layer network are updated based on the goodness of performance. The specific process includes the following steps: Step S41: The feedforward network model consists of two fully connected layers. The initial weight matrix of each layer adopts a random normal distribution, and the activation function is the Sigmoid function, which is used to realize the non-linear mapping of features. Step S42: First, train the first layer of the network. The output vector of the first layer is obtained by inputting sample pairs into the network and is denoted as L. The goodness of L is calculated as the optimization target for model training. Step S43: During training, maximize the goodness value of positive samples and minimize the goodness value of negative samples, and use a stochastic gradient descent optimizer to update the network weight parameters.
6. The fault diagnosis method based on feedforward network and sample combination strategy according to claim 5, characterized in that, In step five, the output L of the first layer is used as the input of the second layer to obtain the output P of the second layer, and the goodness is calculated. The parameters of the second layer are then updated based on the goodness. The specific process includes the following steps: Step S51: Input L into the second layer network to obtain the output vector of the second layer, denoted as P. Calculate the goodness of P as the optimization target for training the second layer. Step S52: During training, maximize the goodness value of positive samples and minimize the goodness value of negative samples, and use a stochastic gradient descent optimizer to update the network weight parameters.
7. The fault diagnosis method based on feedforward network and sample combination strategy according to claim 6, characterized in that, In step six, the bearing inspection data to be tested is sequentially paired with existing data to form sample pairs, which are then input into the trained feedforward network. The output of the second layer network is used to calculate the goodness of fit, and the bearing fault type is determined based on the goodness of fit. The specific process includes the following steps: Step S61: Construct input sample pairs: Perform discrete Fourier transform and normalization on the samples to be detected, combine them with the data of each category of historical data, input each sample pair into the discrete network, and the model outputs the goodness value of each sample pair. Step S62, Fault Category Determination: Identify the sample pair with the highest excellence value. The corresponding category of the historical data contained in this sample pair is the fault diagnosis result.