An industrial equipment fault diagnosis method and system under a complementary label constraint

Abstract

This invention discloses a method and system for fault diagnosis of industrial equipment under complementary label constraints, belonging to the field of industrial intelligent operation and maintenance and fault prediction health management technology. The method includes: collecting and preprocessing equipment vibration signals; constructing a dual-head deep network based on one-dimensional residuals as a complementary label learning model; generating a joint probabilistic loss function using complementary labels based on maximum likelihood estimation, enabling the model to learn fault features from "negative" labels; and achieving high-precision fault classification using the trained model. This invention applies the complementary label learning paradigm in weakly supervised learning to industrial equipment fault diagnosis, achieving a diagnostic accuracy of over 90% on multiple standard bearing datasets such as CWRU and MFPT. This method significantly reduces the data labeling threshold and effectively solves the pain points of scarce fault samples and difficult labeling in industrial scenarios, possessing significant practical application value.

Description

Technical Field

[0001] This invention relates to a method and system for fault diagnosis of industrial equipment under supplementary label constraints, belonging to the field of underwater optical imaging color restoration and image enhancement. Background Technology

[0002] Rotating machinery (such as bearings, gears, and motors) is a core component of industrial systems, and its health directly affects production safety and efficiency. Deep learning-based fault diagnosis methods based on vibration signals have shown great potential. However, the success of these methods heavily relies on a large amount of precisely labeled fault data. In real-world industrial scenarios, acquiring fault samples is costly, and assigning precise category labels to these samples (such as "inner ring fault 0.1mm" or "outer ring fault 0.2mm") requires deep involvement from domain experts, which is time-consuming and labor-intensive, becoming a bottleneck for technology implementation.

[0003] To reduce the cost of data annotation, researchers have proposed weakly supervised learning, which aims to train models using imperfect and imprecise labels. Complementary Label Learning (CLL) is an emerging paradigm that uses "complementary labels" to indicate that a sample does not belong to one or more categories. For example, an experienced engineer may not be able to immediately determine what specific fault a vibration signal corresponds to, but can relatively easily judge that "this is not a normal state" or "this is not an outer ring fault." This negative knowledge is easier to acquire. However, existing research on CLL has largely focused on computer vision. How to design deep learning models and loss functions that can effectively learn from one-dimensional vibration signals and fully utilize the information from the complementary label set is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] Purpose of the invention: To address the problem that fault diagnosis of industrial equipment in the prior art relies heavily on accurate labeled data and has high labeling costs, this invention provides a fault diagnosis method and system under supplementary labeling constraints, which can extract true fault features from negative information and significantly reduce the data labeling threshold.

[0005] Technical solution: Firstly, a method for fault diagnosis of industrial equipment under supplementary label constraints, comprising the following steps:

[0006] Acquire the raw vibration signals generated by the target industrial equipment during operation;

[0007] The original vibration signal is preprocessed, including at least signal slicing and normalization operations, to form a standardized sample sequence that conforms to the model input format.

[0008] A fault diagnosis model based on deep learning is constructed. The model includes a deep feature extraction network and two output layers. The deep feature extraction network is used to extract deep feature vectors, the first output layer is used to fit the probability distribution of fault states, and the second output layer is used to fit the probability distribution of complementary label generation.

[0009] A training dataset is constructed and a model training process is executed. The training dataset includes a sample sequence and a complement label set for each sample. The complement label of a sample is the label of the fault state to which the sample does not belong. The training dataset is input into the fault diagnosis model and a deep feature vector is obtained through a deep feature extraction network. The deep feature vector is passed through two output layers to output the corresponding probability distributions. The joint probabilistic loss function value based on maximum likelihood estimation is calculated based on the two probability distributions. The backpropagation algorithm is used to update all trainable parameters in the fault diagnosis model by minimizing the loss.

[0010] The vibration signal of the industrial equipment to be diagnosed is processed in the same way to form a sample, which is then input into the fault diagnosis model that has been trained. The probability of the sample belonging to each fault state is obtained, and the fault state with the highest probability value is determined as the final diagnosis result of the vibration signal to be diagnosed.

[0011] Furthermore, signal slicing involves dividing a continuous long sequence of vibration signals into fixed-length sliding windows, with all signal data points contained within a sliding window constituting a sample; the normalization operation includes performing zero-mean, unit-variance standardization on each sample individually.

[0012] Furthermore, the deep feature extraction network employs a one-dimensional convolutional neural network with a residual learning structure, specifically comprising an input convolutional block, four sequentially connected residual stages, and a global pooling layer. The input convolutional block consists of a one-dimensional convolutional layer, a batch normalization layer, and a modified linear unit activation function connected sequentially. Each residual stage is composed of multiple basic residual blocks stacked with the same output feature map size. Each basic residual block contains two one-dimensional convolutional sub-layers and an identity shortcut connection spanning the two one-dimensional convolutional sub-layers. The global pooling layer is used to aggregate the feature maps output by the residual stages in the time dimension, outputting a fixed-dimensional deep feature vector.

[0013] Furthermore, the first output layer is a fully connected layer, whose input is a depth feature vector and whose output dimension is the number of fault states. The output is denoted as , indicating sample Prior probability distributions for each fault state These are model parameters;

[0014] The second output layer is a fully connected layer, whose input is a depth feature vector and whose output dimension is the square of the total number of fault states. The output is denoted as The output was reshaped to matrix form Used to fit the true class as Under the condition, category Selected as a sample The conditional probability of the complement label. For model parameters, The second output layer is used to calculate the matrix. The weight parameter for each element value.

[0015] Furthermore, the joint probabilistic loss function based on maximum likelihood estimation is specifically as follows:

[0016]

[0017] in, Denotes the joint probabilistic loss function. Represents the training dataset. Indicates sample The complement of the label set, It represents the set of all fault states, including normal states and various fault categories; Indicates sample The set of candidate true categories; The prior probability of the output of the first output layer. Indicates sample Fault category The prior probability distribution, Indicates sample Fault category The prior probability distribution; This represents the reshaped matrix elements of the second output layer, used to fit the true class. Under the condition, category The conditional probability of being selected as the complement label. and All of these are model parameters.

[0018] Furthermore, the target industrial equipment is rotating machinery, including rolling bearings, gearboxes, and electric motors; the fault status covers various fault modes, including the health status of the equipment, damage to inner ring components, damage to outer ring components, and damage to rolling elements, as well as different levels of damage severity under each fault mode.

[0019] Secondly, an industrial equipment fault diagnosis system under supplementary label constraints includes:

[0020] The signal acquisition module is used to acquire the raw vibration signals generated by the target industrial equipment during operation.

[0021] The signal preprocessing module allows the user to preprocess the original vibration signal. The preprocessing includes at least signal slicing and normalization operations to form a standardized sample sequence that conforms to the model input format.

[0022] The model building module is used to build a fault diagnosis model based on deep learning. The model includes a deep feature extraction network and two output layers. The deep feature extraction network is used to extract deep feature vectors, the first output layer is used to fit the probability distribution of the fault state, and the second output layer is used to fit the probability distribution of the supplementary label generation.

[0023] The model training module is used to construct the training dataset and execute the model training process. The training dataset includes sample sequences and a complement label set for each sample. The complement label of a sample is the label of the fault state to which the sample does not belong. The training dataset is input into the fault diagnosis model and a deep feature vector is obtained through a deep feature extraction network. The deep feature vector is passed through two output layers to output the corresponding probability distributions. The joint probabilistic loss function value based on maximum likelihood estimation is calculated based on the two probability distributions. The backpropagation algorithm is used to update all trainable parameters in the fault diagnosis model by minimizing the loss.

[0024] The fault diagnosis module is used to input samples formed by the vibration signals of the industrial equipment to be diagnosed after the same preprocessing into the trained fault diagnosis model, obtain the probability of the sample belonging to each fault state, and determine the fault state corresponding to the highest probability value as the final diagnosis result of the vibration signal to be diagnosed.

[0025] Thirdly, the present invention also provides an electronic device comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the industrial equipment fault diagnosis method under supplementary label constraints as described in the first aspect of the present invention.

[0026] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the industrial equipment fault diagnosis method under supplementary label constraints as described in the first aspect of the present invention.

[0027] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the industrial equipment fault diagnosis method under supplementary label constraints as described in the first aspect of the present invention.

[0028] Beneficial effects:

[0029] (1) This invention applies complementary label learning to fault diagnosis of industrial equipment. It uses an easily obtainable set of “negative” labels (i.e., the complementary label set of samples that do not belong to a certain class) for training, which greatly reduces the dependence on accurate fault state labels and has strong engineering practicality.

[0030] (2) The present invention has excellent diagnostic performance. The deep feature extraction network obtains deep feature vectors based on the powerful feature extraction capability of one-dimensional ResNet. Based on the principle of maximum likelihood estimation, the joint probabilistic loss function value of the two output layers is constructed by minimizing the negative log-likelihood function to realize the learning loss under the supplementary label constraint. The fault diagnosis accuracy is significantly higher than that of existing methods on multiple public bearing datasets such as CWRU, MFPT, and PU, which verifies its effectiveness.

[0031] (3) The present invention has strong generalization and versatility. The method does not depend on specific equipment or signal processing techniques. The model architecture is universal and can be easily transferred to different types of rotating machinery fault diagnosis tasks. The training is stable and efficient. It adopts an end-to-end training method, combined with an adaptive optimizer and learning rate scheduling, so the model converges quickly and the training process is stable. Attached Figure Description

[0032] Figure 1 The above is a flowchart of the industrial equipment fault diagnosis method under supplementary label constraints provided in the embodiments of the present invention.

[0033] Figure 2 This is a schematic diagram of the network structure of the label learning model of the present invention.

[0034] Figure 3 This is a comparison of the accuracy curves of the present invention with those of two complementary label learning methods on the CWRU dataset.

[0035] Figure 4 This paper compares the accuracy curves of the present invention with those of two methods based on complementary label learning on the PU dataset.

[0036] Figure 5 This paper compares the accuracy curves of the present invention with those of two methods based on complementary label learning on the MFPT dataset. Detailed Implementation

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and should not be used to limit the scope of protection of the present invention.

[0038] This invention proposes a framework for fault diagnosis of industrial equipment using complementary label information in weakly supervised learning problems. The main idea is to achieve high-precision fault classification by utilizing only the "negative" supervisory information (i.e., complementary labels) provided by equipment vibration signal samples, combined with a deep neural network and a specially designed complementary label loss function. Figure 1 As shown, the present invention provides a method for fault diagnosis of industrial equipment under supplementary label constraints, comprising the following steps:

[0039] Step S1: Data preparation and preprocessing.

[0040] This step aims to convert the raw vibration signal into high-quality, normalized data suitable for training deep learning models, and specifically includes the following sub-steps:

[0041] Data Acquisition: For the target rotating machinery (such as rolling bearings), IEPE-compliant accelerometers are installed at key measuring points to continuously acquire vibration acceleration signals of the equipment under different health states (normal, inner race fault, outer race fault, rolling element fault, etc.) and different operating conditions at a sampling frequency of not less than 12.8 kHz. During the model training phase, training is performed using samples and their corresponding complementary label sets; the actual fault state labels of the samples are only used for performance evaluation after model training is completed and do not participate in the model training process.

[0042] Signal slicing and sample generation: Continuous long-sequence vibration signals are segmented into fixed-length sliding windows to prevent information loss and construct a sufficient sample set. A sample refers to all signal data points contained within a sliding window, i.e., a complete signal sequence. The preferred window length is 1024 data points, and the sliding step size can be set to 512 points.

[0043] Data preprocessing and augmentation: Each sample was individually standardized to zero mean and unit variance. The calculation formula is as follows: ,in and These represent the mean and standard deviation of the sample, respectively. During the training phase, for fault states with insufficient sample size, simple data augmentation techniques such as temporal scaling and small random shifts are used, or random overlapping slices are taken from other samples of the same class to alleviate class imbalance.

[0044] Step S2: Construct a complementary label learning model for fault diagnosis of industrial equipment.

[0045] This step constructs a one-dimensional deep convolutional neural network model designed for fault diagnosis of industrial equipment, referred to in this invention as the fault diagnosis model or the complementary label learning model. The model includes a deep feature extraction network and two output layers. The deep feature extraction network extracts deep feature vectors, the first output layer fits the probability distribution of the fault state, and the second output layer fits the probability distribution of the complementary label generation. (Refer to...) Figure 2 The specific components are as follows:

[0046] Feature extraction network: Contains one input convolutional block, four residual stages, and a global pooling layer, outputting a deep feature vector. The input convolutional block consists of a one-dimensional convolutional layer, a batch normalization layer, and a modified linear unit activation function (ReLU) connected sequentially. The weights of this layer are initialized through random distribution during the initial training phase. In this embodiment, the input convolutional block consists of a one-dimensional convolutional layer with a kernel size of 7, a stride of 2, and padding of 3, a batch normalization layer (BN), and a modified linear unit activation function (ReLU) connected sequentially. Each residual stage consists of multiple stacked basic residual blocks with the same output feature map size. Each basic residual block contains two one-dimensional convolutional sub-layers and an identity shortcut connection spanning the two one-dimensional convolutional sub-layers. In this embodiment, each residual stage consists of two stacked basic residual blocks with the same output feature map size. To ensure consistent feature map size, each basic residual block uses a convolutional layer with a kernel size of 3 and padding of 1, along with a stride of 1, ensuring dimensionality matching during residual connections. The global pooling layer aggregates the feature maps output by the residual stages along the time dimension, outputting a fixed-dimensional depth feature vector. In this embodiment of the invention, the global pooling layer outputs a 512-dimensional depth feature vector.

[0047] The first output layer is a fully connected layer, also known as the classifier output layer. , represented as , For the sample size, The total number of fault states is represented by a deep feature vector. The output dimension is the number of fault states. The first output layer outputs the prior probability of the fault, denoted as... , indicating sample Prior probability distributions for each fault state These are the model parameters.

[0048] The second output layer is a fully connected layer, also known as the model-aided output layer. , represented as Its input is a deep feature vector. Output dimension The second output layer outputs the conditional probability with a supplementary label, denoted as... The output was reshaped to matrix form .in For model parameters, Specifically refers to the second output layer used to calculate the matrix. The matrix contains specific weight parameters for each element value and is used to fit the true class. Under the condition, category Selected as a sample The conditional probability of the complement label.

[0049] Step S3: Model training based on the complement label set.

[0050] The core of this step lies in utilizing the complement mark set. and joint probabilistic loss function To optimize this invention, the sample is recorded. The actual fault status label is ,in , It is a set of all fault states (including normal states and various fault categories). This represents the total number of fault states. (Note: The last part is a typo and can be left as is.) For the sample The set of complementary labels, where the complementary labels are the samples. If it is not a fault state, then The complement set is: and .

[0051] The specific steps for calculating the label-addition learning loss function in this invention are as follows:

[0052] (S31) Define the relationship between probability output and conditional probability: Let the classifier output of the fault diagnosis deep learning model be... Let the model's auxiliary output be... The Sigmoid function is used to define a given true class as... At that time, category Selected as a sample The conditional probabilities of the complement label are as follows:

[0053]

[0054] Correspondingly, categories The conditional probability of not belonging to the complement label set (i.e., belonging to the candidate true class set) is:

[0055]

[0056] (S32) Calculate marginal probabilities based on the law of total probability: using classifier output As prior probability Calculate any category Belongs to the current sample complement set The probability of:

[0057]

[0058] and categories The probability of belonging to the non-complementary label set:

[0059]

[0060] (S33) Constructing the joint probabilistic loss function: Based on the principle of maximum likelihood estimation, the final label learning loss function is constructed by minimizing the negative log-likelihood function. :

[0061]

[0062] in, This represents the training dataset, which includes sample sequences and a complement label set for each sample. Represents the set of all fault states. Indicates sample The complement set of symbols, Indicates sample The set of candidate true categories.

[0063] The joint probabilistic loss function proposed in this invention characterizes the complement label learning process from the perspective of probabilistic generation modeling. It treats the true fault state of the sample as a latent variable and considers the complement label to be the observation result randomly generated from the true fault state through a complement label generation mechanism. Unlike existing methods that simply treat the complement label as a negative label or assume it is uniformly generated, this invention introduces a second output layer to explicitly fit "when the true fault category is..." Under these conditions, fault state The conditional probability of being selected as a complement label is used to characterize the class bias and imbalance that may exist in the complement label generation process. Based on the above modeling idea, this invention adopts the maximum likelihood estimation principle to jointly model the prior probability of the fault state obtained from the first output layer and the conditional probability obtained from the second output layer. By minimizing the negative log-likelihood of the complement label observation probability, a unified joint probabilistic loss function is constructed. This loss function is not a simple superposition of multiple loss terms, but rather a holistic optimization of the model parameters under the same probabilistic framework, so that the deep feature extraction network is simultaneously guided by the complement label constraint and the true class distribution constraint. Through the above design, the negativity contained in the complement label set is reduced. Information is transformed into probabilistic constraints on the candidate true fault state space. During training, the model can gradually reduce the probability of impossible fault states and focus on more probable fault states. In industrial equipment fault diagnosis tasks, this joint probabilistic loss function enables the model to learn more discriminative vibration feature representations, allowing different fault states to form a clearer distinguishable structure in the feature space. Experimental results show that although the model fluctuates somewhat in the early stages of training due to the complexity of random parameter initialization and joint modeling, as training progresses, this loss function provides more stable and effective gradient information, thereby accelerating model convergence and significantly improving the final fault diagnosis accuracy.

[0064] This invention uses the backpropagation algorithm to calculate the loss function. The gradients of each parameter of the model are calculated, and the model parameters are updated using stochastic gradient descent or its variants (such as the Adam optimizer). and The iteration continues until the model converges.

[0065] Step S4: Use the trained model to perform online fault diagnosis.

[0066] The vibration signals of the industrial equipment to be diagnosed, after undergoing the same preprocessing, are input into a pre-trained deep learning model for fault diagnosis; the probability of the sample belonging to each fault state is then obtained. The fault state with the highest probability value is determined as the final diagnostic result of the vibration signal to be diagnosed.

[0067] The target industrial equipment to which the method of this invention is applicable is rotating machinery, including but not limited to rolling bearings, gearboxes and electric motors; the fault status covers the health status of the equipment, damage to inner ring components, damage to outer ring components, damage to rolling elements and other fault modes, as well as different levels of damage severity under each fault mode.

[0068] To verify the effectiveness of the method of this invention, experiments were conducted on three publicly available standard bearing fault datasets: CWRU, MFPT, and PU. Two complementary label learning methods published in recent years in the journal Neural Networks were selected for experimental comparison. One method was MVSLDCLL (Tang L, Yan P, Tian Y, et al. Self-adaptive label discovery and multi-view fusion for complementary label learning[J]. Neural Networks, 2025, 181:106763.), published in Neural Networks, Volume 181. The other method was ComCo (Tian Y, Tang L, Liu G, et al. ComCo: Complementary supervised contrastive learning for complementary label learning[J]. Neural Networks, 2024, 169:44-56.), published in Neural Networks, Volume 169. The results are shown in Tables 1, 2, and 3. Figure 3 , Figure 4 , Figure 5 As shown.

[0069] Table 1. Accuracy comparison of various methods on the CWRU dataset (unit: %)

[0070] Epoch MVSLDCLL ComCo This method 1 67.50 45.84 30.52 10 66.00 76.36 92.53 20 66.50 78.90 100 30 66.50 83.51 100 40 67.00 73.83 100 50 66.00 90.26 100 60 65.00 88.70 100 70 64.50 87.92 100 80 63.50 91.69 100 90 63.00 91.04 100 100 63.00 91.82 100

[0071] Table 2. Accuracy comparison of various methods on the PU dataset (unit: %)

[0072] Epoch MVSLDCLL ComCo This method 1 64.44 66.92 70.00 10 64.17 78.84 82.89 20 63.89 81.90 87.19 30 63.06 84.67 87.44 40 62.22 84.57 91.78 50 62.22 85.83 92.39 60 61.39 86.11 92.67 70 60.28 85.23 92.69 80 59.17 86.09 93.06 90 57.78 85.61 92.75 100 56.94 85.34 93.00

[0073] Table 3. Accuracy comparison of various methods on the MFPT dataset (unit: %)

[0074] Epoch MVSLDCLL ComCo This method 1 47.69 32.21 27.51 10 48.04 61.77 64.80 20 49.11 56.64 65.73 30 49.11 60.00 67.37 40 49.11 61.86 69.93 50 49.11 61.12 79.49 60 49.47 62.98 80.89 70 48.75 62.70 86.95 80 48.04 62.61 90.21 90 48.40 63.87 88.81 100 47.69 60.19 89.98

[0075] As shown in Tables 1-3, the accuracy of the method in this invention is lower than that of existing methods in the first epoch, but it quickly surpasses them thereafter. Analysis reveals that MVSLDLL is a method based on handcrafted features (RMS, kurtosis, spectral energy, etc.) and mathematical optimization, directly obtaining a global approximate solution by solving matrix equations. Therefore, its "first epoch" is essentially the result after the calculation, starting close to its peak. The ComCo method employs a warm-up strategy and incorporates a contrastive learning framework, typically imposing strong constraints on the feature space, resulting in slightly smoother initial convergence. The deep neural network convolutional kernel weights used in this method are randomly initialized in the first epoch, leading to initially chaotic feature extraction. Furthermore, the complex loss function, coupled with the inclusion of two networks in the core, causes oscillations in the early training stages. However, the features extracted by the ResNet network improve later, and thanks to the complex and powerful loss function correction, the accuracy rapidly increases.

[0076] Combination Figures 3-5 On the CWRU bearing dataset, which contains 10 states, the classification accuracy of the test set using the method of this invention is higher than that of existing methods.

[0077] MFPT bearing dataset: contains 9 states, and the final test accuracy reaches a high diagnostic precision.

[0078] The Paderborn University (PU) bearing dataset contains nine real-world damage states, which are more challenging, yet the method described in this invention still achieves stable and good results.

[0079] The above experimental results fully demonstrate that the fault diagnosis method under supplementary label constraints provided by this invention can achieve high-precision diagnostic performance comparable to fully supervised learning under extremely low-cost supervision that only requires weak "negativity" labels, effectively solving the major problem of accurate labeling of fault samples in industrial practice.

[0080] Based on the same technical concept as the method embodiments, the present invention also provides an industrial equipment fault diagnosis system under supplementary label constraints, comprising:

[0081] The signal acquisition module is used to acquire the raw vibration signals generated by the target industrial equipment during operation.

[0082] The signal preprocessing module allows the user to preprocess the original vibration signal. The preprocessing includes at least signal slicing and normalization operations to form a standardized sample sequence that conforms to the model input format.

[0083] The model building module is used to build a fault diagnosis model based on deep learning. The model includes a deep feature extraction network and two output layers. The deep feature extraction network is used to extract deep feature vectors, the first output layer is used to fit the probability distribution of the fault state, and the second output layer is used to fit the probability distribution of the supplementary label generation.

[0084] The model training module is used to construct the training dataset and execute the model training process. The training dataset includes sample sequences and a complement label set for each sample. The complement label of a sample is the label of the fault state to which the sample does not belong. The training dataset is input into the fault diagnosis model and a deep feature vector is obtained through a deep feature extraction network. The deep feature vector is passed through two output layers to output the corresponding probability distributions. The joint probabilistic loss function value based on maximum likelihood estimation is calculated based on the two probability distributions. The backpropagation algorithm is used to update all trainable parameters in the fault diagnosis model by minimizing the loss.

[0085] The fault diagnosis module is used to input samples formed by the vibration signals of the industrial equipment to be diagnosed after the same preprocessing into the trained fault diagnosis model, obtain the probability of the sample belonging to each fault state, and determine the fault state corresponding to the highest probability value as the final diagnosis result of the vibration signal to be diagnosed.

[0086] It should be understood that the industrial equipment fault diagnosis system under supplementary label constraints in the embodiments of the present invention can realize all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above embodiments, which will not be repeated here.

[0087] The present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the industrial equipment fault diagnosis method under complementary label constraints as described above.

[0088] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the industrial equipment fault diagnosis method under complementary label constraints as described above.

[0089] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, computer devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

Claims

1. A method for fault diagnosis of industrial equipment under complementary label constraints, characterized in that, Includes the following steps: Acquire the raw vibration signals generated by the target industrial equipment during operation; The original vibration signal is preprocessed, including at least signal slicing and normalization operations, to form a standardized sample sequence that conforms to the model input format. A fault diagnosis model based on deep learning is constructed. The model includes a deep feature extraction network and two output layers. The deep feature extraction network is used to extract deep feature vectors, the first output layer is used to fit the probability distribution of fault states, and the second output layer is used to fit the probability distribution of complementary label generation. Construct a training dataset and execute the model training process. The training dataset includes a sample sequence and a complement label set for each sample. The complement label of a sample is the fault state label to which the sample does not belong. The training dataset is input into the fault diagnosis model, and a deep feature vector is obtained through a deep feature extraction network. This deep feature vector is then passed through two output layers, each outputting a probability distribution. A joint probabilistic loss function based on maximum likelihood estimation is calculated based on these two probability distributions. Using the backpropagation algorithm, all trainable parameters in the fault diagnosis model are updated by minimizing this loss. Specifically, the joint probabilistic loss function based on maximum likelihood estimation is as follows: ； in, Denotes the joint probabilistic loss function. Represents the training dataset. Indicates sample The complement of the label set, It represents the set of all fault states, including normal states and various fault categories; Indicates sample The set of candidate true categories; The prior probability of the output of the first output layer. Indicates sample Fault category The prior probability distribution, Indicates sample Fault category The prior probability distribution; This represents the reshaped matrix elements of the second output layer, used to fit the true class. Under the condition, category The conditional probability of being selected as the complement label. For model parameters, The second output layer is used to calculate the matrix. Weight parameters for each element value; The vibration signal of the industrial equipment to be diagnosed is processed in the same way to form a sample, which is then input into the fault diagnosis model that has been trained. The probability of the sample belonging to each fault state is obtained, and the fault state with the highest probability value is determined as the final diagnosis result of the vibration signal to be diagnosed.

2. The method according to claim 1, characterized in that, Signal slicing involves dividing a continuous long sequence of vibration signals into fixed-length sliding windows, with all signal data points contained within a sliding window considered as a sample; normalization operations include standardizing each sample individually with zero mean and unit variance.

3. The method according to claim 1, characterized in that, The deep feature extraction network employs a one-dimensional convolutional neural network with a residual learning structure, specifically comprising an input convolutional block, four sequentially connected residual stages, and a global pooling layer. The input convolutional block consists of a one-dimensional convolutional layer, a batch normalization layer, and a modified linear unit activation function connected sequentially. Each residual stage is composed of multiple basic residual blocks with the same output feature map size stacked together. Each basic residual block contains two one-dimensional convolutional sub-layers and an identity shortcut connection spanning the two one-dimensional convolutional sub-layers. The global pooling layer is used to aggregate the feature maps output by the residual stages in the time dimension, outputting a fixed-dimensional deep feature vector.

4. The method according to claim 1, characterized in that, The first output layer is a fully connected layer, whose input is a depth feature vector and whose output dimension is the number of fault states. The output is denoted as , indicating sample Prior probability distributions for each fault state These are model parameters; The second output layer is a fully connected layer, whose input is a depth feature vector and whose output dimension is the square of the total number of fault states. The output is denoted as The output was reshaped to matrix form Used to fit the true class as Under the condition, category Selected as a sample The conditional probability of the complement label. For model parameters, The second output layer is used to calculate the matrix. The weight parameter for each element value.

5. The method according to claim 1, characterized in that, The target industrial equipment is rotating machinery, including rolling bearings, gearboxes, and electric motors; the fault status covers various fault modes, including the health status of the equipment, damage to inner ring components, damage to outer ring components, and damage to rolling elements, as well as different levels of damage severity under each fault mode.

6. A fault diagnosis system for industrial equipment under complementary label constraints, characterized in that, include: The signal acquisition module is used to acquire the raw vibration signals generated by the target industrial equipment during operation. The signal preprocessing module allows the user to preprocess the original vibration signal. The preprocessing includes at least signal slicing and normalization operations to form a standardized sample sequence that conforms to the model input format. The model building module is used to build a fault diagnosis model based on deep learning. The model includes a deep feature extraction network and two output layers. The deep feature extraction network is used to extract deep feature vectors, the first output layer is used to fit the probability distribution of the fault state, and the second output layer is used to fit the probability distribution of the supplementary label generation. The model training module is used to construct the training dataset and execute the model training process. The training dataset includes sample sequences and a set of complementary labels for each sample. The complementary labels of a sample are the fault state labels to which the sample does not belong. The training dataset is input into the fault diagnosis model, and a deep feature vector is obtained through a deep feature extraction network. This deep feature vector is then passed through two output layers, each outputting a probability distribution. A joint probabilistic loss function based on maximum likelihood estimation is calculated based on these two probability distributions. Using the backpropagation algorithm, all trainable parameters in the fault diagnosis model are updated by minimizing this loss. Specifically, the joint probabilistic loss function based on maximum likelihood estimation is as follows: ； in, Denotes the joint probabilistic loss function. Represents the training dataset. Indicates sample The complement of the label set, It represents the set of all fault states, including normal states and various fault categories; Indicates sample The set of candidate true categories; The prior probability of the output of the first output layer. Indicates sample Fault category The prior probability distribution, Indicates sample Fault category The prior probability distribution; This represents the reshaped matrix elements of the second output layer, used to fit the true class. Under the condition, category The conditional probability of being selected as the complement label. For model parameters, The second output layer is used to calculate the matrix. Weight parameters for each element value; The fault diagnosis module is used to input samples formed by the vibration signals of the industrial equipment to be diagnosed after the same preprocessing into the trained fault diagnosis model, obtain the probability of the sample belonging to each fault state, and determine the fault state corresponding to the highest probability value as the final diagnosis result of the vibration signal to be diagnosed.

7. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the industrial equipment fault diagnosis method under the supplementary label constraint as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the industrial equipment fault diagnosis method under the supplementary label constraint as described in any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the industrial equipment fault diagnosis method under the supplementary label constraint as described in any one of claims 1-5.

Images