A method, system, and medium for diagnosing mechanical composite faults based on multi-model collaborative decision-making.
The mechanical composite fault diagnosis method based on multi-model collaborative decision-making utilizes multiple lightweight models to make collaborative decisions on equipment time-series signals, solving the problems of low accuracy and high computational resource consumption of single deep learning models in mechanical fault diagnosis, and achieving efficient, accurate fault diagnosis and real-time performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-30
- Publication Date
- 2026-04-03
AI Technical Summary
In existing technologies, single deep learning models have low accuracy and high computational resource consumption in mechanical fault diagnosis, making them difficult to promote in actual industrial diagnostic tasks. Furthermore, traditional signal processing methods rely on engineers' experience and lack universal applicability.
A mechanical composite fault diagnosis method based on multi-model collaborative decision-making is adopted. By expanding the equipment time-series signal into multiple samples through sliding window sampling, multiple lightweight fault detection models are used for collaborative decision-making. The final fault category is determined by combining the fault detection results and label processing.
It significantly improves the accuracy of mechanical fault diagnosis, reduces computing resource consumption, achieves efficient end-to-end fault diagnosis, and adapts to the real-time and convenient nature of industrial environments.
Smart Images
Figure CN119416025B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of mechanical fault diagnosis, and more specifically, relates to a mechanical composite fault diagnosis method, system and medium based on multi-model collaborative decision-making. Background Technology
[0002] In modern industrial production, mechanical equipment plays a vital role in fields such as machinery manufacturing, aerospace, and transportation. Its stable operation is directly related to production safety and efficiency. During actual operation, modern intelligent equipment is prone to various malfunctions due to factors such as material fatigue, improper assembly, and lubrication failure. These malfunctions include bearing damage, gear wear, and rotor imbalance. These problems not only reduce equipment performance and cause economic losses, but can even lead to major safety accidents.
[0003] In actual industrial production environments, although fault mechanism-based diagnosis provides a deep understanding of the root causes of faults in theory, it faces challenges such as high model complexity and difficulty in modeling. Traditional signal processing-based diagnostic methods rely on engineers' experience in feature extraction and selection, threshold setting, etc., lack universally applicable criteria, and are difficult to predict the trend of fault development.
[0004] With the further development of sensor and big data technologies, the amount of data samples extracted in industry is gradually increasing, making data-driven fault diagnosis methods promising. Deep learning, with its automatic feature learning and extraction capabilities, powerful expressive power, and adaptive capabilities, has been widely used in mechanical fault diagnosis. Through multi-layered nonlinear transformations of neural networks, deep learning possesses high expressive power, enabling it to handle complex fault signals and patterns, thereby improving diagnostic accuracy. Furthermore, deep learning models operate on an end-to-end learning model, realizing the transformation from raw input data to fault diagnosis results. This greatly simplifies the complexity of multiple independent processing steps in traditional diagnostic systems, improving the efficiency of the diagnostic process.
[0005] In existing technologies for fault diagnosis using deep learning, the accuracy of fault diagnosis is low in some cases due to the randomness of decisions made by a single deep learning model. In addition, the large amount of computational resources required for training deep learning models makes it difficult to promote in practical industrial diagnostic tasks. Summary of the Invention
[0006] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a mechanical composite fault diagnosis method, system and medium based on multi-model collaborative decision-making, the purpose of which is to improve the accuracy of equipment fault diagnosis.
[0007] To achieve the above objectives, according to a first aspect of the present invention, a mechanical composite fault diagnosis method based on multi-model collaborative decision-making is proposed, comprising the following steps:
[0008] The timing signal of the device to be diagnosed is expanded into Y samples by sliding window sampling;
[0009] Y samples are input into X pre-trained fault detection models. Each fault detection model performs fault detection on each sample, resulting in a total of X×Y fault detection results.
[0010] The fault detection results are converted into fault labels. The fault detection results reflect the probability that the sample belongs to each type of fault, and the fault labels reflect the fault category.
[0011] For any fault detection model: Y input samples correspond to Y fault labels. Based on the number of samples corresponding to the fault labels, the model fault labels are determined; and the model fault labels of all fault detection models are obtained accordingly.
[0012] The two model fault labels that appear most frequently among all model fault labels are selected as decision labels, and the number of fault detection models and the number of samples corresponding to the two decision labels are determined respectively.
[0013] If the difference in the number of fault detection models corresponding to two decision labels is not less than a preset threshold, then the decision label with more corresponding fault detection models is used as the final fault label; otherwise, the decision label with more corresponding samples is used as the final fault label; the fault category reflected by the final fault label is the equipment fault category.
[0014] As a further preferred option, the fault detection results are converted into fault labels, including:
[0015] For any fault detection result, the probability of the sample belonging to each type of fault is determined based on the fault detection result. If the probability exceeds 50%, the corresponding fault category is retained to form a fault label, which can be a single fault or a compound fault.
[0016] As a further preferred approach, the model fault labels are determined based on the number of samples corresponding to the fault labels, including:
[0017] Each of the Y fault labels is divided into individual faults, and the number of corresponding samples is counted. All individual faults with a sample count greater than Y / 2 are selected and retained. If no individual fault has a sample count greater than Y / 2, the individual fault with the largest sample count is retained. The retained individual faults are combined to construct the model fault labels.
[0018] As a further preferred option, after obtaining all model fault labels, if no more than half of the model fault labels are the same type of composite fault, then the detected composite fault is regarded as several single faults, and the sample corresponding to the composite fault is used as the sample of the corresponding single fault, and then the subsequent decision label is determined.
[0019] As a further preferred option, there are a total of 7 fault detection models, which are constructed based on TFSCL, MSRCN, CFSPT, CFSPT_SD, HFPT, MSMCNN_SD, and Props2, respectively.
[0020] As a further preferred option, X fault detection models are pre-trained, including: acquiring time-series signals for each fault category, and obtaining corresponding samples through sliding window sampling to construct a training set, and training each fault detection model using the training set.
[0021] As a further preferred option, when determining the final fault label, if the number of fault detection models and the number of samples corresponding to the two decision labels are the same, or the number of samples corresponding to the two decision labels is less than Y / 2, then the final fault label is determined manually.
[0022] As a further preferred embodiment, the timing signal is a vibration, sound, or rotational speed signal of the device.
[0023] According to a second aspect of the present invention, a mechanical composite fault diagnosis system based on multi-model collaborative decision-making is provided, comprising a processor for executing the above-described mechanical composite fault diagnosis method based on multi-model collaborative decision-making.
[0024] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described mechanical composite fault diagnosis method based on multi-model collaborative decision-making.
[0025] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages:
[0026] 1. This invention designs a multi-sample, multi-model collaborative fault diagnosis method. By converting equipment time-series signals into multi-sample batch decision-making and integration, it effectively improves the comprehensiveness and complementarity of information perception and significantly enhances the generalization performance of the model. At the same time, unlike the traditional single-model strategy, this invention adopts multiple detection models to collaboratively fuse decision-making, which significantly reduces the risk of overfitting a single model to specific data and improves the accuracy of equipment fault diagnosis.
[0027] 2. This invention employs several lightweight detection models, and the training process does not involve any target domain information. After training using historical data, the models can directly perform diagnosis without any domain adaptation or transfer learning processes to perform secondary processing on the test data, thus exhibiting high real-time performance. Furthermore, it can be easily embedded into a distributed training framework, allowing the use of multiple training devices to accelerate the overall training time and reduce computational resource consumption.
[0028] 3. This invention is based on data-driven deep learning-based fault diagnosis technology, which can achieve end-to-end diagnosis without the need for complex expert knowledge, effectively solving problems such as high turnover of expert personnel and difficulty in accumulating in-depth business knowledge in industry verticals. Attached Figure Description
[0029] Figure 1 This is a schematic diagram of the mechanical composite fault diagnosis method based on multi-model collaborative decision-making according to an embodiment of the present invention;
[0030] Figure 2 This is a schematic diagram of sliding window sampling according to an embodiment of the present invention;
[0031] Figure 3 This is a schematic diagram of fault detection results and fault labels in an embodiment of the present invention;
[0032] Figure 4 This is a schematic diagram of the multi-model collaborative decision-making process in an embodiment of the present invention;
[0033] Figure 5 This is a flowchart illustrating the multi-model voting rules in an embodiment of the present invention.
[0034] Figure 6 (a) and (b) are schematic diagrams of the fault decision-making process in an embodiment of the present invention. Detailed Implementation
[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0036] This invention provides a method for diagnosing complex mechanical faults based on multi-model collaborative decision-making, such as... Figure 1 and Figure 4 As shown, it includes the following steps:
[0037] S1. Before diagnosis, pre-train the fault detection model.
[0038] Time-series signals for various fault categories are collected using industrial sensors, and the data is stored in CSV format according to the sensor channels. Time-series signal data from known fault type files are sampled using a sliding window to obtain corresponding samples, thus constructing a training set. X fault detection models are then trained using this training set. Specifically, the time-series signals are vibration, sound, or rotational speed signals from the equipment.
[0039] Furthermore, this embodiment includes seven fault detection models, comprising four CNN models and three Transformer models, specifically TFSCL, MSRCN, CFSPT, CFSPT_SD, HFPT, MSMCNN_SD, and Props2. These seven models are smaller in size, faster in computation, more efficient, and more accurate than mainstream models.
[0040] S2. Obtain the timing signal of the device to be diagnosed, and expand the timing signal of the device to be diagnosed into Y samples by sliding window sampling; input the Y samples into X pre-trained fault detection models respectively, and each fault detection model performs fault detection on each sample, resulting in a total of X×Y fault detection results.
[0041] Specifically, the sliding window sampling in S1 and S2 is as follows: Figure 2 As shown, the length of each sample in the sample unit is the step size, which is the length of the interval between the previous sample and the next sample. The overlapping part represents the tail of the previous sample and the head of the next sample. This method can make the most of the limited dataset to generate more samples.
[0042] S3. Based on the X×Y fault detection results obtained from X fault detection models, determine the fault type of the equipment according to the established voting rules.
[0043] Specifically, the voting rules of S3 mainly include four parts: preliminary processing of decision data, failure mode statistics, preliminary failure decision, and final failure decision.
[0044] S31. Preliminary processing of decision data: The fault detection results are converted into fault labels. The fault detection results reflect the probability that the sample belongs to each type of fault, and the fault labels reflect the fault category.
[0045] The specific rules for converting fault detection results into fault labels are as follows: For any fault detection result, based on the probability that the sample reflected by the fault detection result belongs to a certain type of fault, if the probability exceeds 50%, the corresponding fault category is retained, thereby forming a fault label; the fault label can be a single fault or a compound fault.
[0046] The above process can be performed using a composite fault decoupler based on the Sigmoid function, such as... Figure 3As shown, the deep learning model represents the fault detection model, and the different shaped icons output by the model represent different faults. The "P" after the icon represents the probability of the corresponding fault occurring. The output of the composite fault decoupling classifier is the fault category determined by the model. Figure 3 The output is a composite fault that includes two types of faults.
[0047] S32. Fault Mode Statistics: For any fault detection model, there are Y fault labels corresponding to the Y input samples. Based on the number of samples corresponding to the fault labels, the model fault labels are determined; accordingly, the model fault labels corresponding to X fault detection models are determined.
[0048] Furthermore, based on the number of samples corresponding to the fault labels, the model fault labels are determined, including: splitting the fault categories corresponding to the Y fault labels into single faults and counting the corresponding sample counts; filtering and retaining all single faults with a sample count greater than Y / 2; if no single fault has a sample count greater than Y / 2, then retaining the single fault with the largest sample count; and combining the retained single faults to construct model fault labels.
[0049] S33. Preliminary Fault Decision: Select the two model fault labels that appear most frequently from all model fault labels, i.e., decision labels, and determine the number of fault detection models and the number of samples corresponding to these two decision labels.
[0050] Furthermore, after obtaining all model fault labels, if no more than half of the model fault labels are the same type of composite fault, the detected composite fault is regarded as several corresponding single faults, and the sample corresponding to the composite fault is used as a sample of several single faults. These samples are then included in the statistical analysis of the relevant single fault samples, summed, and then the preliminary fault decision is performed to determine the decision label. This avoids the situation where the single fault decision is abnormal due to the statistical analysis of composite faults.
[0051] S34. Final Fault Decision: If the difference in the number of fault detection models corresponding to the two model fault labels is not less than a preset threshold, the fault label of the model with more corresponding fault detection models is used as the final fault label; otherwise, the fault label of the model with more corresponding samples is used as the final fault label. The fault category reflected by the final fault label is the equipment fault category. The final decision result is recorded and converted into the required CSV format.
[0052] To make the present invention clearer, specific embodiments are described below.
[0053] First, the timing signal of the device to be diagnosed is expanded to 200 samples using sliding window sampling. These 200 samples are then input into seven models, with each model performing 200 fault diagnoses, resulting in a total of 1400 fault detection results. Then, as... Figure 5As shown, the fault type of the equipment is determined according to the voting rules, including:
[0054] S31. Perform data statistics on 200 fault detection results for each model according to a threshold of 0.5, and convert the probability into fault labels. For example, the continuous numerical probability output by the model is [0.98,0,0.86], which is converted into a label (1,3), indicating that it is a composite fault of fault 1 and fault 3; 7 models yield 7 CSV files.
[0055] S32. For any fault detection model, among 200 fault labels, the composite faults are split into single faults, and their corresponding sample counts are also assigned to the single faults. Then, the samples are statistically analyzed together with other single faults, and all single faults with a sample count greater than 100 are selected (if none are found, the single fault with the largest sample count is selected) and combined into model fault labels.
[0056] S33. Select the two most frequently occurring model fault labels from the seven model fault labels as decision labels, and determine the number of fault detection models and the number of samples corresponding to each decision label. For composite fault labels, the number of samples corresponding to each individual fault may be inconsistent; the average value can be used for subsequent statistics.
[0057] After the above process, the following is formed: Figure 6 The document shown represents the diagnostic result of a time-series signal for each line. Figure 6 The results of multiple diagnostic processes described above are displayed. The blue part represents the model failure label and the corresponding number of decisions (i.e., the corresponding number of samples) obtained from 200 decision statistics of each model; the green part combines the two decision labels of the results of 7 models; the pink part represents the number of models that predict the decision label category, with a maximum value of 7; the purple part represents the number of times that decision label is predicted in all models, with a maximum value of 1400.
[0058] S34. Combine the final votes from different models with the total number of samples to make a collaborative decision and determine the final fault label:
[0059] (1) For two decision labels, if the model voting difference is not less than 2, the statistical model number shall prevail (i.e., voting round with Max=7).
[0060] For example: to Figure 6 In the first row, 6 models are identified as [8,9], and 1 model is identified as [8,9,13]. Since 6-1 is greater than 2, the final fault label is determined to be [8,9].
[0061] (2) For two decision labels, if the model voting difference is less than 2, the number of valid decision samples shall be used (i.e., voting round with Max = 1400).
[0062] For example: If 3 models identify the decision label as A and 3 models identify the decision label as B, then compare the total number of samples for A and B. If 3A(332) > 3B(256), then the decision is A. If 2 models identify the decision label as A and 3 models identify the decision label as B, then compare the total number of samples for A and B. If 2A(320) > 3B(300), then the decision is A.
[0063] (3) In extreme cases where none of the above mechanisms can resolve the issue, human intervention can be used for decision-making.
[0064] For example, if the number of votes and the number of samples are the same, or if all decisions have fewer than 100, they are considered invalid decisions. For example, 1A(80) = 1B(80), which may mean that the type is an unknown fault category that has never appeared before.
[0065] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A mechanical composite fault diagnosis method based on multi-model collaborative decision-making, characterized in that, Includes the following steps: The timing signal of the device to be diagnosed is expanded into Y samples by sliding window sampling; Y samples are input into X pre-trained fault detection models. Each fault detection model performs fault detection on each sample, resulting in a total of X samples. Y fault detection results; The fault detection results are converted into fault labels. The fault detection results reflect the probability that the sample belongs to each type of fault, and the fault labels reflect the fault category. For any fault detection model: Y input samples correspond to Y fault labels. Based on the number of samples corresponding to the fault labels, the model fault labels are determined; and the model fault labels of all fault detection models are obtained accordingly. The two model fault labels that appear most frequently among all model fault labels are selected as decision labels, and the number of fault detection models and the number of samples corresponding to the two decision labels are determined respectively. If the difference between the number of fault detection models corresponding to two decision labels is not less than a preset threshold, then the decision label with more corresponding fault detection models will be used as the final fault label. Otherwise, the decision label with the larger number of corresponding samples is used as the final fault label; the fault category reflected by the final fault label is the equipment fault category.
2. The mechanical composite fault diagnosis method based on multi-model collaborative decision-making as described in claim 1, characterized in that, The fault detection results are converted into fault labels, including: For any fault detection result, the fault detection result is converted into a fault label according to the probability that the sample reflected by the fault detection result belongs to each type of fault: if the probability exceeds 50%, the corresponding fault category is retained, thus forming a fault label, which can be a single fault or a compound fault.
3. The mechanical composite fault diagnosis method based on multi-model collaborative decision-making as described in claim 2, characterized in that, Based on the number of samples corresponding to the fault labels, the model fault labels are determined, including: Each of the Y fault labels is divided into individual faults, and the number of corresponding samples is counted. All individual faults with a sample count greater than Y / 2 are selected and retained. If no individual fault has a sample count greater than Y / 2, the individual fault with the largest sample count is retained. The retained individual faults are combined to construct the model fault labels.
4. The mechanical composite fault diagnosis method based on multi-model collaborative decision-making as described in claim 2, characterized in that, After obtaining all model fault labels, if no more than half of the model fault labels are the same composite fault, the detected composite fault is regarded as several single faults, and the sample corresponding to the composite fault is used as the sample of the corresponding single fault, and then the subsequent decision label is determined.
5. The mechanical composite fault diagnosis method based on multi-model collaborative decision-making as described in claim 1, characterized in that, There are a total of 7 fault detection models, which are built based on TFSCL, MSRCN, CFSPT, CFSPT_SD, HFPT, MSMCNN_SD, and Props2, respectively.
6. The mechanical composite fault diagnosis method based on multi-model collaborative decision-making as described in claim 1, characterized in that, Pre-train X fault detection models, including: acquiring time-series signals for each fault category and obtaining corresponding samples through sliding window sampling to construct a training set, and training each fault detection model using the training set.
7. The mechanical composite fault diagnosis method based on multi-model collaborative decision-making as described in claim 1, characterized in that, When determining the final fault label, if the number of fault detection models and the number of samples corresponding to the two decision labels are the same, or the number of samples corresponding to the two decision labels is less than Y / 2, then the final fault label is determined manually.
8. The mechanical composite fault diagnosis method based on multi-model collaborative decision-making as described in any one of claims 1-7, characterized in that, The timing signal is the vibration, sound, or rotation speed signal of the device.
9. A mechanical composite fault diagnosis system based on multi-model collaborative decision-making, characterized in that, Includes a processor for executing the mechanical composite fault diagnosis method based on multi-model collaborative decision-making as described in any one of claims 1-8.
10. 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 mechanical composite fault diagnosis method based on multi-model collaborative decision-making as described in any one of claims 1-8.