Hybrid evidence-based expert method and system for open-set fault diagnosis of rotating machinery

Abstract

The application discloses a rotating machinery open set fault diagnosis method and system based on a hybrid evidence expert, and the method first collects vibration data of rotating machinery parts for pretreatment, and constructs an open set fault diagnosis sample set.Secondly, a binary evidence expert model is constructed for each known fault category, and independent training is carried out.Then, a non-target fault evidence suppression supervision label is constructed, and each binary evidence expert model is fine-tuned to obtain a trained hybrid evidence expert model.Finally, a sample to be diagnosed is input into the trained hybrid evidence expert model, and each binary evidence expert model output is obtained to determine corresponding category probability and uncertainty, and normal state recognition, known fault classification and unknown fault rejection are carried out, and a diagnosis result is output.The application realizes normal state recognition, known fault classification and unknown fault rejection, and improves the open set fault diagnosis performance of the rotating machinery.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis, and in particular to a method and system for diagnosing open set faults in rotating machinery based on hybrid evidence experts. Background Technology

[0002] Rotating machinery is widely used in energy, transportation, manufacturing, and industrial equipment, serving as crucial equipment for power transmission and mechanical motion. Key components such as bearings, gears, and drive shafts typically operate under conditions of high speed, heavy load, variable speed, or complex loads, and their operational status directly impacts the safety and reliability of the equipment. Therefore, timely and accurate fault diagnosis of key components in rotating machinery is of great significance for ensuring safe equipment operation and reducing maintenance costs.

[0003] Vibration signals reflect the dynamic characteristics of rotating machinery during operation and are commonly used monitoring signals in rotating machinery fault diagnosis. Traditional fault diagnosis methods typically rely on manual feature extraction and empirical discrimination rules, and their diagnostic effectiveness is easily affected by changes in operating conditions, noise interference, and feature selection methods. With the accumulation of industrial monitoring data and the improvement of computing power, deep learning-based fault diagnosis methods are gradually being applied to the field of rotating machinery fault diagnosis. These methods can automatically learn fault characteristics from vibration signals and achieve the identification of normal states or fault categories.

[0004] However, existing deep learning-based fault diagnosis methods for rotating machinery are mostly based on the closed-set diagnosis assumption, which assumes that the health states or fault categories included in the training phase can cover all categories that may occur in the testing phase. In real industrial scenarios, rotating machinery has complex structures and diverse fault modes. Some fault types are characterized by suddenness, rarity, or difficulty in pre-collection, making it difficult for training data to cover all fault types that may occur during equipment operation. When unknown fault categories appear in the testing phase that were not seen in the training phase, the closed-set diagnosis model may still classify them into a known category, leading to unreliable diagnostic results.

[0005] Therefore, fault diagnosis of rotating machinery needs to be expanded from traditional closed-set classification to open-set fault diagnosis scenarios. Open-set fault diagnosis not only needs to identify normal states and known fault categories, but also needs to have the ability to reject unknown fault samples. To achieve the rejection of unknown faults, the diagnostic model cannot only output the classification results of known categories, but also needs to characterize the uncertainty of the diagnostic results. Evidence-based learning methods can transform the model output into evidence supporting different categories, and calculate uncertainty while obtaining the category probabilities, providing a basis for rejecting unknown faults.

[0006] However, rotating machinery faults are complex, and the vibration signal characteristics corresponding to different fault categories differ. When a single diagnostic model learns multiple category discrimination relationships simultaneously, it may be difficult to fully represent the local discrimination features of different target fault categories. Therefore, how to construct an open-set fault diagnosis method for rotating machinery that can combine the local discrimination ability, evidence expression ability, uncertainty estimation ability, and fusion reasoning mechanism of multiple evidence expert models, and simultaneously achieve normal state identification, known fault classification, and unknown fault rejection, is a problem that needs to be solved in this field. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for open-set fault diagnosis of rotating machinery based on hybrid evidence experts. This invention solves the problems of existing deep learning-based rotating machinery fault diagnosis methods, which struggle to simultaneously meet the requirements of known fault classification and unknown fault rejection, and the difficulty of a single diagnostic model adequately representing the local discriminative features of different target fault categories. To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0008] In a first aspect, the present invention provides a method for diagnosing open-set faults in rotating machinery based on hybrid evidence experts, comprising the following steps:

[0009] S1. Collect vibration data of key components of rotating machinery under different health conditions.

[0010] S2. Preprocess the collected vibration data to construct an open-set fault diagnosis sample set.

[0011] S3. Construct a binary evidence expert model for each known fault category in the open set fault diagnosis sample set. Each binary evidence expert model uses the normal category and a corresponding known fault category as the discrimination object and is used to output the evidence information of the input sample relative to the normal state and the corresponding known fault category.

[0012] S4. Select normal samples and known fault category samples corresponding to each binary evidence expert model from the open set of fault diagnosis samples, and train each binary evidence expert model independently.

[0013] S5. Set non-target fault samples as evidence suppression objects and construct non-target fault evidence suppression supervision labels, wherein the non-target fault samples are other known fault category samples in the open set fault diagnosis sample set other than the corresponding known fault category.

[0014] S6. Based on the joint fine-tuning supervision label containing the non-target fault evidence suppression supervision label, perform joint fine-tuning on each binary evidence expert model to suppress the evidence output of non-target fault samples in normal and target fault states, and obtain the trained hybrid evidence expert model.

[0015] S7. Input the sample to be diagnosed into the trained hybrid evidence expert model to obtain the normal state evidence and target fault state evidence output by each binary evidence expert model, and construct a model based on the normal state evidence and target fault state evidence. The parameters determine the corresponding class probability and uncertainty.

[0016] S8. Based on the category probabilities and uncertainties output by each binary evidence expert model, the normal state identification, known fault classification, and unknown fault rejection are performed through the hybrid evidence expert fusion reasoning module, and the diagnostic results are output.

[0017] Secondly, the present invention also provides a rotating machinery open-set fault diagnosis system based on hybrid evidence experts, comprising the following modules:

[0018] The open-set fault diagnosis sample module is used to collect vibration data of rotating machinery components under different health conditions, perform preprocessing, and construct an open-set fault diagnosis sample set.

[0019] The binary evidence expert module is used to construct a binary evidence expert model for each known fault category in the open set of fault diagnosis sample sets, and to train each binary evidence expert model independently.

[0020] The hybrid evidence expert module is used to set non-target fault samples as evidence suppression objects, construct non-target fault evidence suppression supervision labels, and jointly fine-tune each binary evidence expert model to obtain the trained hybrid evidence expert model.

[0021] The fault diagnosis output module is used to input the sample to be diagnosed into the trained hybrid evidence expert model, obtain the output of each binary evidence expert model, and construct... The parameters determine the corresponding category probabilities and uncertainties. Based on the category probabilities and uncertainties output by each binary evidence expert model, the system performs normal state identification, known fault classification, and unknown fault rejection, and outputs diagnostic results.

[0022] The present invention has the following beneficial effects:

[0023] 1. Enhance the ability to discriminate evidence for known fault categories. This invention constructs a binary evidence expert model for each known fault category, enabling different binary evidence expert models to learn the evidence discrimination relationship between normal state and corresponding known fault state, thereby improving the targeted diagnostic ability for different known fault categories, by using the normal category and the corresponding known fault category as discrimination objects respectively.

[0024] 2. Improve the uncertainty representation capability of non-target fault samples. This invention uses samples of known fault categories other than their corresponding known fault categories as non-target fault samples, and constructs a joint fine-tuning supervision label containing a non-target fault evidence suppression supervision label. This label is then used to jointly fine-tune each binary evidence expert model to suppress the evidence output of non-target fault samples in normal and target fault states. This ensures that the corresponding binary evidence expert model maintains a high degree of uncertainty for non-target fault samples, providing a basis for rejecting unknown faults.

[0025] 3. Improve the performance of open-set fault diagnosis for rotating machinery. This invention proposes a hybrid evidence expert fusion reasoning module, which constructs a system based on the normal state evidence and target fault state evidence output by each binary evidence expert model. The parameters are determined, and the corresponding category probabilities and uncertainties are identified. The hybrid evidence expert fusion reasoning module integrates the category probabilities and uncertainties output by each binary evidence expert model to make a fusion decision, thereby realizing normal state identification, known fault classification, and unknown fault rejection, thus improving the open set fault diagnosis performance of rotating machinery. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments will be briefly described below. The features and advantages of the present invention can be more clearly understood by referring to the accompanying drawings, which are illustrative and should not be construed as limiting the present invention in any way. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. Wherein:

[0027] Figure 1 This is a flowchart of the open set fault diagnosis method for rotating machinery based on hybrid evidence experts according to the present invention;

[0028] Figure 2 This is a schematic diagram of the hybrid evidence expert model structure of the present invention;

[0029] Figure 3 This is a flowchart of the training process for the hybrid evidence expert model of this invention;

[0030] Figure 4 This is a flowchart of the diagnostic reasoning process based on mixed evidence experts in this invention;

[0031] Figure 5 This is a comparison chart of the open-set fault diagnosis confusion matrices of the CNN-EDL model, the LSTM-EDL model, and the method of this invention.

[0032] Figure 6 This is a comparison chart of the uncertainty distributions of known category samples and unknown fault samples using the CNN-EDL model, the LSTM-EDL model, and the method of this invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] In this embodiment of the invention, a method and system for open-set fault diagnosis of rotating machinery based on hybrid evidence experts are proposed. The flowchart of the method is as follows: Figure 1 As shown, it includes the following steps:

[0035] S1. Collect vibration data of key components of rotating machinery under different health conditions.

[0036] S2. Preprocess the collected vibration data to construct an open-set fault diagnosis sample set.

[0037] S3. Construct a binary evidence expert model for each known fault category in the open set fault diagnosis sample set. Each binary evidence expert model uses the normal category and a corresponding known fault category as the discrimination object and is used to output the evidence information of the input sample relative to the normal state and the corresponding known fault category.

[0038] S4. Select normal samples and known fault category samples corresponding to each binary evidence expert model from the open set of fault diagnosis samples, and train each binary evidence expert model independently.

[0039] S5. Set non-target fault samples as evidence suppression objects and construct non-target fault evidence suppression supervision labels, wherein the non-target fault samples are other known fault category samples in the open set fault diagnosis sample set other than the corresponding known fault category.

[0040] S6. Based on the joint fine-tuning supervision label containing the non-target fault evidence suppression supervision label, perform joint fine-tuning on each binary evidence expert model to suppress the evidence output of non-target fault samples in normal and target fault states, and obtain the trained hybrid evidence expert model.

[0041] S7. Input the sample to be diagnosed into the trained hybrid evidence expert model to obtain the normal state evidence and target fault state evidence output by each binary evidence expert model, and construct a model based on the normal state evidence and target fault state evidence. The parameters determine the corresponding class probability and uncertainty.

[0042] S8. Based on the category probabilities and uncertainties output by each binary evidence expert model, the normal state identification, known fault classification, and unknown fault rejection are performed through the hybrid evidence expert fusion reasoning module, and the diagnostic results are output.

[0043] In a preferred embodiment of the present invention, the method mainly includes two processes: hybrid evidence expert model training and diagnostic reasoning. The hybrid evidence expert model training process and the diagnostic reasoning process are respectively as follows: Figure 3 and Figure 4 As shown.

[0044] In the training process of the hybrid evidence expert model, the collected vibration data of rotating machinery is preprocessed to construct an open set fault diagnosis sample set; a corresponding binary evidence expert model is constructed for each known fault category, and each binary evidence expert model is independently trained and jointly fine-tuned in sequence to obtain the trained hybrid evidence expert model.

[0045] In the diagnostic reasoning process, the sample to be diagnosed is preprocessed in the same way as in the training phase, then input into the trained hybrid evidence expert model to obtain the category probability and uncertainty output by each binary evidence expert model, and the diagnostic result is output through the hybrid evidence expert fusion reasoning module.

[0046] Furthermore, step S1 is specifically implemented as follows:

[0047] In this embodiment of the invention, the vibration data can be acquired by an acceleration sensor located near a key component of the rotating machinery, which includes at least one of a bearing, gear, drive shaft, or planetary gearbox. Alternatively, the vibration data can be vibration signal data from a publicly available dataset of rotating machinery faults.

[0048] Furthermore, step S2 is specifically implemented as follows:

[0049] Select a target sensor channel from the vibration data to obtain the vibration signal corresponding to the target sensor channel; divide the vibration signal into equal length segments to obtain multiple equal-length vibration signal samples; construct features from the equal-length vibration signal samples to obtain corresponding diagnostic samples; and construct an open-set fault diagnosis sample set based on the diagnostic samples.

[0050] Specifically, in this embodiment, the feature construction includes: performing wavelet packet decomposition on the equal-length vibration signal samples to extract the coefficients of each sub-band; arranging the coefficients of each sub-band in frequency order to construct a two-dimensional time-frequency feature matrix, and using the two-dimensional time-frequency feature matrix as the diagnostic sample.

[0051] The sample set is composed of the two-dimensional time-frequency feature matrix and its corresponding category label. ,in, Indicates the first The two-dimensional time-frequency feature matrix corresponding to each vibration signal sample Indicates the first The category label corresponding to each vibration signal sample Indicates the number of samples.

[0052] Furthermore, the open-set fault diagnosis sample set includes a training sample set for model training and samples to be diagnosed input during the diagnosis phase; wherein, the training sample set includes normal samples and samples of known fault categories, but does not include samples of unknown fault categories to be rejected. During the diagnosis phase, the sample to be diagnosed is input into the trained hybrid evidence expert model; if it does not meet the normal state determination condition and any known fault category determination condition, then an unknown fault diagnosis result is output.

[0053] In this embodiment of the invention, let the known category set be... The set of unknown categories is ,in:

[0054]

[0055] in Indicates the normal category. They represent Known fault categories, The number of known fault categories. The set of unknown categories. This includes unknown fault categories that did not participate in model training during the training phase, and satisfies the following:

[0056]

[0057] Therefore, only the sample set is used in the training phase. Category tags belong to The mixed evidence expert model is trained using normal samples and known faulty samples, and the category labels belong to... Unknown fault samples are not included in the model parameter update.

[0058] Furthermore, step S3 is specifically implemented as follows:

[0059] like Figure 2As shown, a binary evidence expert model is constructed for each known fault category. Let the number of known fault categories be . Regarding the first Construct the first known fault category A binary evidence expert model, in which... The first A binary evidence expert model with normal category and the first The known fault categories are used as the discrimination objects, and are used to output the normal state evidence corresponding to the input sample and the first fault category. Evidence of the target fault state corresponding to a known fault category; The aforementioned binary evidence expert models together constitute a hybrid evidence expert model.

[0060] In this embodiment, the first The binary evidence expert model includes a feature extraction part and a binary output part. The feature extraction part is used to extract fault features from the input diagnostic sample; the binary output part is used to output binary results corresponding to the normal state and the target fault state based on the fault features. The binary results are then subjected to non-negative mapping to obtain evidence of the normal state and evidence of the target fault state.

[0061] Furthermore, step S4 is specifically implemented as follows:

[0062] For the first When a binary evidence expert model is trained independently, normal samples and the second binary evidence expert model are selected from the known class samples during the training phase. By analyzing fault samples corresponding to known fault categories, a binary training subset is obtained:

[0063]

[0064] in, Indicates the first The binary training subset used by the binary evidence expert model during the independent training phase. Sourced from sample set During the training phase, samples of known categories are available. Indicates the normal category. Indicates the first Known fault categories.

[0065] During the independent training phase, the labels for normal samples are set to... , will the The fault sample labels corresponding to the known fault categories are set as follows: Then the first The binary supervision labels corresponding to each binary evidence expert model are:

[0066]

[0067] in, Indicates the first The binary evidence expert model is for the first Independent training supervision labels were set for each sample.

[0068] Input samples from the binary training subset into the first... A binary evidence expert model was used to obtain the original output results:

[0069]

[0070] in, Indicates the first A binary evidence expert model, This indicates the output result under normal conditions. This indicates the output result of the target fault status.

[0071] The original output is nonnegatively mapped to obtain the evidence vector:

[0072]

[0073] in, Evidence indicating a normal state Evidence indicating the target's fault state This represents the activation function used to map the original output to non-negative evidence.

[0074] Constructed based on the evidence vector parameter:

[0075]

[0076] in, Indicates normal state parameter, Indicates the target fault state parameter, Represents a two-dimensional vector of all 1s Vector addition is performed on corresponding elements.

[0077] Based on the binary supervision label and Parameters, using binary evidence learning loss for the first... A binary evidence expert model is trained, and the binary evidence learning loss is expressed as:

[0078]

[0079] in, Indicates the first Binary evidence learning loss of a binary evidence expert model during the independent training phase. This represents the number of samples in the binary training subset. This represents the expected mean square error term. Indicates the first The annealing coefficient corresponding to each training round Indicates the use of calculation divergence regularization term parameter, express Divergence regularization term.

[0080] The expected mean square error term is expressed as:

[0081]

[0082] in, Corresponding to the normal state, Corresponding to the target fault state, Indicates binary supervision label The One portion, express parameter The One portion, The expected mean squared error term includes the category label and... Error terms between distribution expectations and Distribution variance term.

[0083] The calculation divergence regularization term The parameters are represented as follows:

[0084]

[0085] in, This indicates element-wise multiplication.

[0086] The The divergence regularization term is expressed as:

[0087]

[0088] in, Indicates by Definite distributed, This indicates that all parameters are equal. of distributed, Indicates two Between distributions Divergence. The aforementioned... The divergence regularization term is used to constrain the evidence output corresponding to non-true categories, so as to reduce the model's overconfidence in non-true categories.

[0089] By minimizing the binary evidence learning loss, the first... The parameters of each binary evidence expert model are configured to enable it to distinguish between normal states and corresponding target fault states, and to determine the corresponding classification label and uncertainty based on the output evidence information. The above independent training process is performed on each known fault category to obtain multiple binary evidence expert models with initial discrimination capabilities.

[0090] Furthermore, step S5 is specifically implemented as follows:

[0091] For the The binary evidence expert model uses normal samples as the discrimination samples for normal states, and the first... The fault samples corresponding to the known fault categories are used as target fault samples, except for the first one. Samples of other known fault categories besides the known fault category are considered as non-target fault samples.

[0092] For the first When performing joint fine-tuning of two binary evidence expert models, the normal sample labels are set to... , will the The fault sample labels corresponding to the known fault categories are set as follows: Set the label of non-target fault samples to The corresponding joint fine-tuning supervision labels are set as follows:

[0093]

[0094] in, Indicates the first The binary evidence expert model is for the first Joint fine-tuning supervision labels for each sample setting Indicates the normal category. Indicates the first Known fault categories, This represents a known set of categories. For the... For a binary evidence expert model, the category label belongs to... And it does not belong to the normal category and the first Fault samples of known fault categories are non-target fault samples of this binary evidence expert model.

[0095] For the non-target fault sample, it belongs neither to the normal category nor to the first category. The target fault category corresponds to each of the binary evidence expert models. Therefore, in the first... In a binary evidence expert model, non-target fault samples should not be learned as normal state samples, nor should they be learned as target fault state samples. Based on this, this embodiment sets the joint fine-tuning supervision label for the non-target fault samples as... This is used to indicate that the sample does not provide positive category supervision at either the normal state output node or the target fault state output node.

[0096] Specifically, for the first The non-target fault sample, following the evidence construction method in step S4, yields the... Evidence vectors corresponding to each binary evidence expert model And constructed from the evidence vector parameter Because the joint fine-tuning supervision label for non-target fault samples is Under the constraint of evidence learning loss, the first A binary evidence expert model provides evidence of the normal state of the non-target fault sample. Evidence of target fault status All are suppressed, causing them to approach 0; correspondingly, parameter Approaching This approaches a state of no prior information, where neither output node has received additional evidence. In this case, the binary evidence expert model neither tends to classify the non-target fault sample as a normal state nor as the first fault. This allows for the identification of known fault categories, thereby improving the expert model's ability to represent uncertainties in non-target fault samples.

[0097] Furthermore, step S6 is specifically implemented as follows:

[0098] After each binary evidence expert model completes independent training, the parameters obtained during the independent training phase are used as the initial parameters for the joint fine-tuning phase. From the sample set... We select normal samples from the training phase and fault samples corresponding to each known fault category to construct a joint fine-tuning training set:

[0099]

[0100] in, This indicates joint fine-tuning of the training set. This represents a set of known categories; the joint fine-tuning training set includes normal samples and fault samples corresponding to multiple known fault categories, but does not include samples of unknown categories.

[0101] During the joint fine-tuning phase, the samples from the joint fine-tuning training set are input into each binary evidence expert model. For the first... A binary evidence expert model, based on the joint fine-tuning supervision labels set in step S5. And the evidence constructed from the output of the binary evidence expert model. parameter Calculate the first Joint fine-tuning of evidence learning loss for two binary evidence expert models:

[0102]

[0103] in, Indicates the first Joint fine-tuning of evidence learning loss for two binary evidence expert models This indicates the number of samples in the joint fine-tuning training set. Indicates the first The annealing coefficient corresponding to each training round Indicates the use of calculation divergence regularization term parameter, express Divergence regularization term.

[0104] Furthermore, due to the joint fine-tuning of the training set This includes normal samples, target fault samples, and non-target fault samples, and the three types of samples participate in the loss calculation according to the joint fine-tuning supervision labels in step S5. Therefore, the first... The joint fine-tuning evidence learning loss of a binary evidence expert model can be composed of normal sample loss, target fault sample loss, and non-target fault evidence suppression loss. The normal sample loss is used to maintain the ability to distinguish normal states, the target fault sample loss is used to maintain the ability to distinguish target fault states, and the non-target fault evidence suppression loss is used to suppress the evidence output of non-target fault samples in both normal and target fault states.

[0105] Based on the joint fine-tuning evidence learning loss of each binary evidence expert model, construct the joint fine-tuning total loss:

[0106]

[0107] in, This indicates the total loss from joint fine-tuning. This indicates the number of binary evidence expert models.

[0108] By minimizing the total loss of the joint fine-tuning and updating the parameters of each binary evidence expert model, a trained hybrid evidence expert model is obtained.

[0109] Furthermore, step S7 is specifically implemented as follows:

[0110] In the diagnosis phase, the sample to be diagnosed is processed into a diagnostic sample X using the same preprocessing method as in the training phase, and this diagnostic sample X is input into the trained hybrid evidence expert model. For the k-th binary evidence expert model, its original output is obtained:

[0111]

[0112] in, This indicates the completion of the training. A binary evidence expert model, This indicates the output result under normal conditions. This indicates the output result of the target fault status.

[0113] Following the evidence construction method in step S4, the original output result is non-negatively mapped to obtain the first... The evidence vector corresponding to each binary evidence expert model:

[0114]

[0115] in, Evidence indicating a normal state Evidence representing the target fault state is used to characterize whether the sample to be diagnosed supports a normal state or a second fault state. The strength of evidence for a known fault category.

[0116] Constructed based on the evidence vector parameter:

[0117]

[0118] in, , .

[0119] make According to the above Parameters determine the first The normal category probability, target fault category probability, and uncertainty corresponding to each binary evidence expert model:

[0120]

[0121]

[0122]

[0123] in, Indicates the first The normal category probability output by a binary evidence expert model. Indicates the first The target fault category probability is output by a binary evidence expert model. Indicates the first The uncertainty of the diagnostic results for each binary evidence expert model is determined. From this, the category probability and uncertainty corresponding to each binary evidence expert model are obtained.

[0124] Furthermore, step S8 is specifically implemented as follows:

[0125] Set the normal judgment threshold Target fault determination threshold Based on the normal category probabilities output by each binary evidence expert model obtained in step S7 Target Fault Category Probability and uncertainty The preliminary diagnostic results and their corresponding uncertainties for the sample to be diagnosed are determined through a hybrid evidence expert fusion reasoning module; the normal judgment threshold is... and target fault determination threshold The preliminary diagnostic results of the sample to be diagnosed and the corresponding uncertainty determination process can be determined based on the normal category probability distribution and the target fault category probability distribution corresponding to the known category samples in the validation set; the preliminary diagnostic results of the sample to be diagnosed and the corresponding uncertainty determination process can be expressed as follows:

[0126]

[0127] in This indicates the preliminary diagnosis results. This indicates the uncertainty corresponding to the preliminary diagnostic results; Indicates the normal category. Indicates the first There are several known fault categories. When a sample to be diagnosed meets neither the normal state determination criteria nor the target fault state determination criteria, it is treated as a fuzzy sample. This represents the auxiliary candidate category with the highest probability of the target fault category. This indicates a value greater than the unknown fault determination threshold. The preset maximum uncertainty value is set to 1. In this embodiment, the preset maximum uncertainty value is set to 1.

[0128] Furthermore, determine the threshold for identifying unknown faults. This is used for unknown fault rejection. The unknown fault determination threshold is... Represented as:

[0129]

[0130] Among them, the uncertainty set The uncertainties are arranged in ascending order. Represents the set of uncertainties The first quartile, or the 25th percentile, Represents the set of uncertainties The third quartile, i.e., the 75th percentile; the uncertainty set The uncertainty of the preliminary diagnosis result is composed of the known category samples in the validation set. The validation set is obtained by dividing the training sample set. The uncertainty of the preliminary diagnosis result is obtained according to the judgment logic of the hybrid evidence expert fusion reasoning module.

[0131] Furthermore, the uncertainty corresponding to the preliminary diagnostic results is... With the unknown fault determination threshold Comparison, final diagnosis results Represented as:

[0132]

[0133] in, Indicates an unknown fault diagnosis result; if If the condition is met, the unknown fault diagnosis result is output; otherwise, the preliminary diagnosis result is output. Thus, the hybrid evidence expert fusion reasoning module can achieve normal state identification, known fault classification, and unknown fault rejection based on the category probabilities and uncertainties output by each binary evidence expert model.

[0134] Secondly, the present invention also provides a rotating machinery open-set fault diagnosis system based on hybrid evidence experts, for implementing the aforementioned rotating machinery open-set fault diagnosis method, including an open-set fault diagnosis sample module, a binary evidence expert module, a hybrid evidence expert module, and a fault diagnosis output module:

[0135] The open-set fault diagnosis sample module is used to collect vibration data of rotating machinery components under different health conditions, perform preprocessing, and construct an open-set fault diagnosis sample set.

[0136] The binary evidence expert module is used to construct a binary evidence expert model for each known fault category in the open set fault diagnosis sample set, and to train each binary evidence expert model independently.

[0137] The hybrid evidence expert module is used to set non-target fault samples as evidence suppression objects, construct non-target fault evidence suppression supervision labels, and jointly fine-tune each binary evidence expert model to obtain the trained hybrid evidence expert model.

[0138] The fault diagnosis output module is used to input the sample to be diagnosed into the trained hybrid evidence expert model, obtain the output of each binary evidence expert model, and construct... The parameters determine the corresponding category probabilities and uncertainties. Based on the category probabilities and uncertainties output by each binary evidence expert model, the system performs normal state identification, known fault classification, and unknown fault rejection, and outputs diagnostic results.

[0139] To verify the effectiveness of the open-set fault diagnosis method for rotating machinery based on hybrid evidence experts disclosed in this invention, this embodiment uses vibration data of a planetary gearbox collected by a Drivetrain Diagnostics Simulator (DDS) for experimental verification. The DDS experimental platform mainly includes a motor, a planetary gearbox, a parallel shaft gearbox, and a magnetic powder brake. An acceleration sensor is installed on the planetary gearbox housing to collect vibration signals under different health conditions.

[0140] In this embodiment, vibration data of a planetary gearbox under variable speed conditions were selected for verification. The speed gradually increased from 20Hz to 38.67Hz, and the sampling frequency was 25600Hz. Each health condition was repeatedly sampled four times. To facilitate sample segmentation, the data from the first 48 seconds of each vibration signal group was selected, and the sensor channel corresponding to the Y-axis direction of the planetary gearbox was selected as the target sensor channel. The vibration signal was divided into segments of equal length, each containing 4096 sampling points. Wavelet packet decomposition was performed on each segment of the vibration signal sample to obtain a 64×64 wavelet packet coefficient matrix as the diagnostic sample.

[0141] In this embodiment, the planetary gearbox vibration data includes 9 health states, and the specific categories and open set division methods are shown in Table 1.

[0142] Table 1 Health Status and Open Set Division of DDS Planetary Gearbox

[0143]

[0144] As shown in Table 1, the known category set in this embodiment is as follows: The set of unknown categories is The training phase only uses The samples in the dataset are used to train various binary evidence expert models. The samples in the test are not used for model training, but only for testing to verify the ability to reject unknown faults.

[0145] Furthermore, 1200 diagnostic samples were constructed for each health state, with 1000 samples used as training samples and 200 samples used as test samples. During the training phase, data augmentation strategies were applied to the training samples to improve the model's adaptability to sampling phase shifts, amplitude fluctuations, and local perturbations; during the testing phase, the aforementioned data augmentation processing was not performed on the diagnostic samples.

[0146] To further verify the open-set fault diagnosis performance of the method of this invention, CNN-EDL and LSTM-EDL models were selected as comparison methods. The CNN-EDL model uses a convolutional network as the feature extraction structure and outputs evidence information corresponding to known categories through evidence learning; the LSTM-EDL model uses a long short-term memory network to extract sample features and outputs evidence information corresponding to known categories through evidence learning. Both of these comparison methods are single multi-class evidence learning models, used to compare with the hybrid evidence expert model described in this invention.

[0147] Furthermore, the known category recognition accuracy OS*, the unknown fault rejection accuracy Uk, the open set comprehensive accuracy OS, and AUROC are used as evaluation indicators for open set fault diagnosis. The calculation formulas for OS*, Uk, and OS are as follows:

[0148]

[0149]

[0150]

[0151] in, This represents the total number of samples with known categories in the test set. This represents the number of known category samples that were correctly identified in the test set; This represents the total number of samples with unknown fault categories in the test set. This represents the number of unknown fault category samples correctly rejected in the test set. AUROC represents the area under the ROC curve, used to characterize the model's ability to distinguish between known category samples and unknown fault samples based on the uncertainty score. The larger the AUROC, the stronger the model's ability to distinguish between known category samples and unknown fault samples.

[0152] The experimental results of different methods in the open set fault diagnosis task of DDS planetary gearbox are shown in Table 2.

[0153] Table 2. Open set fault diagnosis results of different methods

[0154]

[0155] As shown in Table 2, although the CNN-EDL model achieves 100.00% accuracy in known category recognition (OS*), which measures the ability to identify known category samples, its accuracy in rejecting unknown fault samples (Uk) is only 64.00%. This indicates that while single multi-class evidence learning models have certain advantages in known category recognition, they are still prone to classifying unknown fault samples into a known category. The LSTM-EDL model has an unknown fault rejection accuracy (Uk) of 33.00%, and an AUROC of 0.7524, which characterizes the overall ability to distinguish between known category samples and unknown fault samples. This suggests that its ability to reject unknown fault samples and its ability to distinguish between known category samples and unknown fault samples are both weak.

[0156] The method of this invention achieves an unknown fault rejection accuracy (Uk) of 97.50% and an AUROC of 0.9917, both higher than the comparative methods. This indicates that the method of this invention can effectively reject unknown fault samples and can effectively distinguish between known category samples and unknown fault samples based on uncertainty. The known category recognition accuracy (OS*) of the method of this invention is 94.25%, and the open set comprehensive accuracy (OS), used to comprehensively measure the correct recognition of known category samples and the correct rejection of unknown fault samples, reaches 94.61%. The above results show that the method of this invention improves the ability to reject unknown faults while maintaining high known category recognition capability and open set comprehensive diagnostic accuracy.

[0157] To further analyze the specific diagnostic performance of different methods on known category samples and unknown fault samples, open-set fault diagnosis confusion matrices for the CNN-EDL model, LSTM-EDL model, and the method of this invention were plotted. The results are as follows: Figure 5 As shown in the diagram. In this diagram, the vertical axis of the confusion matrix represents the true class of the sample, the horizontal axis represents the predicted class output by the model, and the numerical values ​​in the matrix represent the proportion of samples of the corresponding true class that are diagnosed as the corresponding predicted class. Figure 5 It can be seen that the CNN-EDL model has a high recognition accuracy for samples of known categories, but it can only correctly reject 64.00% of unknown fault samples, with the remaining unknown fault samples mainly being misidentified as normal states. The LSTM-EDL model can only correctly reject 33.00% of unknown fault samples, and some unknown fault samples are misidentified as normal states or known fault categories. In contrast, the method of this invention can correctly reject 97.50% of unknown fault samples, and the number of unknown fault samples misclassified as known categories is significantly reduced, indicating that the method of this invention can more effectively reject unknown fault samples.

[0158] To further analyze the uncertainty representation capabilities of different methods for known category samples and unknown fault samples, the uncertainty distributions of the CNN-EDL model, LSTM-EDL model, and the method of this invention on the test set were statistically analyzed. The results are as follows: Figure 6 As shown in the figure, the horizontal axis represents the sample uncertainty used for unknown fault determination, and the vertical axis represents the probability density of the sample uncertainty; the higher the sample uncertainty, the lower the reliability of the model in diagnosing the sample as a known category. Figure 6 It can be seen that the uncertainty distributions of known class samples and unknown fault samples in both the CNN-EDL and LSTM-EDL models overlap to varying degrees. In contrast, the known class samples in the method of this invention are mainly concentrated in the low uncertainty region, while the unknown fault samples are mainly concentrated in the high uncertainty region, and the uncertainty distributions of the two types of samples are clearly distinguishable. The above results are consistent with Table 2 and... Figure 5 The results are consistent with those shown, indicating that the method of the present invention can improve the uncertainty characterization ability of unknown fault samples and effectively reject unknown fault samples based on uncertainty.

[0159] In summary, the quantitative evaluation results shown in Table 2, Figure 5 The confusion matrix results shown and Figure 6 The uncertainty distribution results shown collectively demonstrate that the open-set fault diagnosis method for rotating machinery based on hybrid evidence experts described in this invention can improve the uncertainty representation and rejection capabilities of unknown fault samples while maintaining a high recognition capability for known categories. This is because this invention constructs a binary evidence expert model for each known fault category and sets non-target fault samples as evidence suppression objects during the joint fine-tuning stage. This allows each binary evidence expert model to suppress evidence output in both normal and target fault states when facing non-target fault samples, thus maintaining a high uncertainty for samples that do not match its discrimination task, providing a basis for rejecting unknown faults. Furthermore, the hybrid evidence expert fusion reasoning module integrates the category probabilities and uncertainties output by each binary evidence expert model to make diagnostic decisions, enabling the model to effectively reject unknown fault samples while recognizing normal states and known fault categories, thereby improving the reliability of open-set fault diagnosis for rotating machinery.

[0160] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for open-set fault diagnosis of rotating machinery based on hybrid evidence experts, characterized in that, Includes the following steps: S1. Collect vibration data of rotating machinery components under different health conditions, perform preprocessing, and construct an open-set fault diagnosis sample set; S2. For each known fault category in the open set fault diagnosis sample set, construct a binary evidence expert model and train each binary evidence expert model independently. S3. Set non-target fault samples as evidence suppression objects, construct non-target fault evidence suppression supervision labels, and jointly fine-tune each binary evidence expert model to obtain the trained hybrid evidence expert model. S4. Input the sample to be diagnosed into the trained hybrid evidence expert model, obtain the output of each binary evidence expert model, and construct... Parameters determine the corresponding class probability and uncertainty; S5. Based on the category probabilities and uncertainties output by each binary evidence expert model, perform normal state identification, known fault classification, and unknown fault rejection, and output the diagnostic results.

2. The open-set fault diagnosis method for rotating machinery based on hybrid evidence experts according to claim 1, characterized in that, The preprocessing includes selecting a target sensor channel from the vibration data to obtain the vibration signal corresponding to the target sensor channel; dividing the vibration signal into equal-length segments to obtain multiple equal-length vibration signal samples; constructing features from the equal-length vibration signal samples to obtain corresponding diagnostic samples; and constructing an open-set fault diagnosis sample set based on the diagnostic samples.

3. The open-set fault diagnosis method for rotating machinery based on hybrid evidence experts according to claim 2, characterized in that, The feature construction includes: performing wavelet packet decomposition on equal-length vibration signal samples to extract the coefficients of each sub-band; arranging the sub-band coefficients in frequency order to construct a two-dimensional time-frequency feature matrix as a diagnostic sample; and forming a sample set by the two-dimensional time-frequency feature matrix and its corresponding category labels. ,in, Indicates the first The two-dimensional time-frequency feature matrix corresponding to each vibration signal sample Indicates the first The category label corresponding to each vibration signal sample Indicates the number of samples.

4. The open-set fault diagnosis method for rotating machinery based on hybrid evidence experts according to claim 3, characterized in that, The open set fault diagnosis sample set includes a training sample set for training and a sample to be diagnosed input during the diagnosis phase; wherein, the training sample set includes normal samples and known fault category samples, but does not include unknown fault category samples to be rejected; during the diagnosis phase, the sample to be diagnosed is input into the trained hybrid evidence expert model, and if it does not meet the normal state judgment condition and any known fault category judgment condition, the unknown fault diagnosis result is output. Let the known set of categories be... The set of unknown categories is , ,in Indicates the normal category. They represent Known fault categories, set of unknown categories This includes unknown fault categories that were not included in the training phase, and The training phase uses only the sample set. Category tags belong to The mixed evidence expert model is trained using normal samples and known faulty samples, and the category labels belong to... Unknown fault samples are not included in parameter updates.

5. The open-set fault diagnosis method for rotating machinery based on hybrid evidence experts according to claim 4, characterized in that, The specific implementation process of constructing a binary evidence expert model for each known fault category in the open-set fault diagnosis sample set is as follows: Regarding the first Construct the first known fault category A binary evidence expert model, in which... ;No. A binary evidence expert model with normal category and the first The known fault categories are used as the discrimination objects, and are used to output the normal state evidence corresponding to the input sample and the first fault category. Evidence of the target fault state corresponding to a known fault category; The aforementioned binary evidence expert models together constitute a hybrid evidence expert model; No. The binary evidence expert model includes a feature extraction part and a binary output part. The feature extraction part is used to extract the fault features of the input diagnostic sample. The binary output part is used to output the corresponding normal state and target fault state based on the fault features. The binary results are then mapped non-negatively to obtain the normal state evidence and the target fault state evidence.

6. The open-set fault diagnosis method for rotating machinery based on hybrid evidence experts according to claim 5, characterized in that, The specific implementation of independently training each binary evidence expert model is as follows: For the first When a binary evidence expert model is trained independently, normal samples and the second binary evidence expert model are selected from the known class samples during the training phase. The fault samples corresponding to the known fault categories are obtained to obtain the first fault sample. The binary training subset used by each binary evidence expert model during the independent training phase. , Sourced from sample set During the training phase, samples of known categories are available. Indicates the normal category. Indicates the first Known fault categories; During the independent training phase, the labels for normal samples are set to... , will the The fault sample labels corresponding to the known fault categories are set as follows: ,use Indicates the first The binary evidence expert model is for the first Independent training supervision labels were set for each sample; Input samples from the binary training subset into the first... A binary evidence expert model is used to obtain the original output results. , This indicates the output result under normal conditions. The output result represents the target fault state; a non-negative mapping is performed on the original output result to obtain the evidence vector. , Evidence indicating a normal state Evidence indicating the target's fault state; Construction based on evidence vectors parameter ,in, Indicates normal state parameter, Indicates the target fault state parameter, This represents a two-dimensional vector consisting entirely of 1s; vector addition is performed element-wise. Based on binary supervision label and Parameters, using binary evidence learning loss for the first... A binary evidence expert model is trained, and the model is updated by minimizing the binary evidence learning loss. The parameters of each binary evidence expert model are configured to enable it to distinguish between normal states and corresponding target fault states, and to determine the corresponding classification label and uncertainty based on the output evidence information. The above independent training process is performed on each known fault category to obtain multiple binary evidence expert models with initial discrimination capabilities.

7. The open-set fault diagnosis method for rotating machinery based on hybrid evidence experts according to claim 6, characterized in that, The specific implementation process of step S3 is as follows: For the The binary evidence expert model uses normal samples as the discrimination samples for normal states, and the first... The fault samples corresponding to the known fault categories are used as target fault samples, except for the first one. Samples of other known fault categories besides the known fault categories are considered as non-target fault samples; For the first When performing joint fine-tuning of two binary evidence expert models, the normal sample labels are set to... , will the The fault sample labels corresponding to the known fault categories are set as follows: Set the label of non-target fault samples to ,use Indicates the first The binary evidence expert model is for the first Joint fine-tuning of supervisory labels for each sample; After each binary evidence expert model completes independent training, the parameters of the binary evidence expert models obtained in the independent training phase are used as the initial parameters for the joint fine-tuning phase, starting from the sample set. We select normal samples from the training phase and fault samples corresponding to each known fault category to construct a joint fine-tuning training set: ; During the joint fine-tuning phase, samples from the joint fine-tuning training set are input into each binary evidence expert model; for the first... A binary evidence expert model, based on the set joint fine-tuning supervision labels. And the evidence constructed from the output of the binary evidence expert model. parameter Calculate the first Joint fine-tuning of evidence learning loss for two binary evidence expert models ;No. The joint fine-tuning evidence learning loss of each binary evidence expert model can be composed of normal sample loss, target fault sample loss and non-target fault evidence suppression loss. Based on the joint fine-tuning evidence learning loss of each binary evidence expert model, the joint fine-tuning total loss is constructed. By minimizing the joint fine-tuning total loss and updating the parameters of each binary evidence expert model, the trained hybrid evidence expert model is obtained.

8. The open-set fault diagnosis method for rotating machinery based on hybrid evidence experts according to claim 7, characterized in that, The specific implementation process of step S4 is as follows: During the diagnosis phase, the sample to be diagnosed is preprocessed into a diagnostic sample X using the same preprocessing method as in the training phase, and then input into the trained hybrid evidence expert model. For the k-th binary evidence expert model, its original output result is obtained. ,in, This indicates the output result under normal conditions. The output indicates the target fault status. Perform a nonnegative mapping on the original output to obtain the first... Evidence vectors corresponding to each binary evidence expert model ,in, Evidence indicating a normal state Evidence representing the target fault state is used to characterize whether the sample to be diagnosed supports a normal state or a second fault state. Strength of evidence for known fault categories; Construction based on evidence vectors parameter ,make ,according to and The ratio determines the first The normal category probability corresponding to a binary evidence expert model ,according to and The ratio determines the first The probability of the target fault category corresponding to a binary evidence expert model According to 2 and The ratio determines the first Uncertainty corresponding to a binary evidence expert model .

9. The open-set fault diagnosis method for rotating machinery based on hybrid evidence experts according to claim 8, characterized in that, The specific implementation process of step S5 is as follows: Set the normal judgment threshold Target fault determination threshold Based on the normal category probability output by each binary evidence expert model Target Fault Category Probability and uncertainty And determine the preliminary diagnostic results and corresponding uncertainties of the sample to be diagnosed; normal judgment threshold. and target fault determination threshold The probability distribution of the normal category and the probability distribution of the target fault category can be determined based on the known category samples in the validation set; the preliminary diagnostic results of the sample to be diagnosed and the process for determining its corresponding uncertainty are expressed as follows: ;in This indicates the preliminary diagnosis results. This indicates the uncertainty corresponding to the preliminary diagnostic results; Indicates the normal category. Indicates the first There are several known fault categories; when neither the normal state determination criteria nor the target fault state determination criteria are met, the sample to be diagnosed is treated as a fuzzy sample. This represents the auxiliary candidate category with the highest probability of the target fault category. This indicates a value greater than the unknown fault determination threshold. The preset maximum uncertainty value; Determine the unknown fault determination threshold for unknown fault rejection. Among them, the set of uncertainties The uncertainties are arranged in ascending order. Represents the set of uncertainties The first quartile, Represents the set of uncertainties The third quartile; uncertainty set The validation set consists of the uncertainty of the preliminary diagnostic results corresponding to the known category samples in the validation set, which is obtained by partitioning the training sample set. Corresponding uncertainty to the preliminary diagnostic results Unknown fault determination threshold If a comparison is made, If the condition is met, the unknown fault diagnosis result will be output; otherwise, the preliminary diagnosis result will be output.

10. A rotating machinery open-set fault diagnosis system based on hybrid evidence experts, used to implement the rotating machinery open-set fault diagnosis method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The open-set fault diagnosis sample module is used to collect vibration data of rotating machinery components under different health conditions, perform preprocessing, and construct an open-set fault diagnosis sample set. The binary evidence expert module is used to construct a binary evidence expert model for each known fault category in the open set of fault diagnosis sample set, and to train each binary evidence expert model independently. The hybrid evidence expert module is used to set non-target fault samples as evidence suppression objects, construct non-target fault evidence suppression supervision labels, and jointly fine-tune each binary evidence expert model to obtain the trained hybrid evidence expert model. The fault diagnosis output module is used to input the sample to be diagnosed into the trained hybrid evidence expert model, obtain the output of each binary evidence expert model, and construct... The parameters determine the corresponding category probabilities and uncertainties. Based on the category probabilities and uncertainties output by each binary evidence expert model, the system performs normal state identification, known fault classification, and unknown fault rejection, and outputs diagnostic results.

Images