A small sample domain generalization fault detection method and system based on wavelet scattering forest

Abstract

The application relates to the technical field of mechanical fault detection, and discloses a small sample domain generalization fault detection method and system based on a wavelet scattering forest, which comprises the following steps: collecting mechanical equipment operation data and performing pretreatment; and constructing a detection model based on a wavelet scattering forest; wherein the detection model comprises a wavelet scattering transformation data enhancement module, a deep stacked forest classification module and a similarity measurement weighting module; the wavelet scattering transformation data enhancement module is used for performing data enhancement on the mechanical equipment operation data, so as to obtain a plurality of sample sets of different scales; the sample sets are input into the deep stacked forest classification module for prediction, so as to obtain a plurality of prediction vectors; the similarity measurement weighting module is used for calculating weight vectors under different scales based on the sample sets, and the prediction vectors are weighted and summed based on the weight vectors to obtain a detection result. Compared with the prior art, the application provides a detection method which has stronger generalization ability and is more widely applicable.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical fault detection, and particularly relates to a small sample domain generalization fault detection method and system based on wavelet scattering forest. BACKGROUND

[0002] Rotating machinery plays a vital role in key industries such as aerospace, robotics, and wind power. Rotating machinery often operates under extremely complex and variable working conditions. During long-term operation, the performance of the equipment is easily damaged due to harsh working environments, resulting in shortened service life and even safety accidents. Therefore, the research on fault diagnosis and predictive maintenance of rotating machinery can not only effectively prolong the service life of the equipment and improve the operation efficiency, but also reduce maintenance costs and provide technical support for the sustainable development of related industries. Deep learning, as a data-based learning method, has been widely applied in the field of fault diagnosis. However, in many practical applications, fault data is difficult to obtain, especially under small sample conditions, the effectiveness of deep learning is greatly limited. By transferring existing knowledge, expanding the training data set, and optimizing feature representation, the precision of fault diagnosis is effectively improved, providing a new idea for solving the problem of data scarcity.

[0003] At present, although methods such as transfer learning and meta-learning have achieved certain results in the field of fault diagnosis, there are still some deficiencies. First, it is difficult to obtain source domain data similar to the target domain in industrial sites, especially under real working conditions, there is a big difference between artificial simulation data and actual fault data, which limits the generalization ability of these methods. Second, existing small sample learning methods mainly focus on improving diagnostic accuracy, and less consider the constraints in actual applications, such as the diversity of equipment and the change of working conditions in industrial environment. This makes the fault diagnosis effect of existing methods under complex working conditions often unsatisfactory, and there is still a lot of room for improvement, which cannot fully meet the needs of industrial production. SUMMARY

[0004] In order to overcome the defects of insufficient generalization ability of the prior art, a small sample domain generalization fault detection method and system based on wavelet scattering forest are proposed.

[0005] To achieve the above technical effects, the technical solutions of the present application are as follows:

[0006] A small sample domain generalization fault detection method based on wavelet scattering forest, comprising the following steps:

[0007] Collecting mechanical equipment operation data and preprocessing;

[0008] A detection model based on wavelet scattering forest is constructed, wherein the detection model comprises a wavelet scattering transformation data enhancement module, a deep stacked forest classification module and a similarity measurement weighting module.

[0009] The mechanical equipment data is input into the detection model, the wavelet scattering transformation data enhancement module is used to perform data enhancement on the mechanical equipment operation data, and a plurality of sample sets of different scales are obtained.

[0010] The sample sets are input into the deep stacked forest classification module for prediction, and a plurality of prediction vectors are obtained.

[0011] The similarity measurement weighting module is used to calculate a weight vector under different scales based on the sample sets, and the prediction vectors are weighted and summed based on the weight vector to obtain a detection result.

[0012] The application further provides a small sample domain generalization fault detection system based on wavelet scattering forest, which comprises:

[0013] A data acquisition module is configured to acquire mechanical equipment operation data and perform preprocessing.

[0014] A wavelet scattering transformation data enhancement module is configured to input the mechanical equipment data into the detection model, perform data enhancement on the mechanical equipment operation data by using the wavelet scattering transformation data enhancement module, and obtain a plurality of sample sets of different scales.

[0015] A deep stacked forest classification module is configured to input the sample sets into the deep stacked forest classification module for prediction, and obtain a plurality of prediction vectors.

[0016] A similarity measurement weighting module is configured to calculate a weight vector under different scales based on the sample sets by using the similarity measurement weighting module, and perform weighted summation on the prediction vectors based on the weight vector to obtain a detection result.

[0017] The application further provides an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the small sample domain generalization fault detection method based on wavelet scattering forest when executing the computer program.

[0018] The application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executable on a processor to implement the small sample domain generalization fault detection method based on wavelet scattering forest.

[0019] Compared with the prior art, the application has the following beneficial effects:

[0020] The application is based on data enhancement of multi-scale wavelet scattering transformation, scattering coefficients extracted by wavelet scattering transformation have the characteristics of translation invariance and deformation stability, and strong generalization features of samples are effectively obtained; secondly, the stability and generalization performance of the model are further enhanced by deep stacking forest, so as to ensure that the rotating machinery fault can be accurately classified and diagnosed under different working conditions; the similarity measurement is used to quantify the class center of each original sample, the weight is reasonably allocated, and the diagnosis result is further optimized. The application can be widely applied to small sample fault diagnosis scenes of various rotating machines, can learn diagnosis knowledge without the aid of auxiliary data set, can use a small number of samples to diagnose faults of different operating conditions and equipment, and provides an efficient and practical solution for solving the actual equipment fault diagnosis problem in industrial production. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A flowchart of a small sample domain generalization fault detection method based on wavelet scattering forest of embodiment 1.

[0022] Figure 2 An architecture diagram of a small sample domain generalization fault detection system based on wavelet scattering forest of embodiment 2.

[0023] Figure 3 A detection flowchart in the detection model based on wavelet scattering forest.

[0024] Figure 4 An architecture diagram of the detection model based on wavelet scattering forest.

[0025] Figure 5 An architecture diagram of the similarity measurement weighting module.

[0026] Figure 6 It is a precision comparison diagram of the method of the application and other fault diagnosis methods.

[0027] Figure 7 It is a precision comparison diagram of the ablation experiment of the method of the application. DETAILED DESCRIPTION

[0028] The drawings are only used for illustrative description and cannot be understood as a limitation on the application;

[0029] For those skilled in the art, it is understandable that some known descriptions in the drawings can be omitted.

[0030] The technical solutions of the application will be further described below in combination with the drawings and embodiments.

[0031] Embodiment 1

[0032] This embodiment proposes a small sample domain generalization fault detection method based on wavelet scattering forest, as shown in Figure 1, which is a flow chart of a small sample domain generalized fault detection method based on wavelet scattering forests according to this embodiment.

[0033] This embodiment proposes a small sample domain generalized fault detection method based on wavelet scattering forest, including the following steps:

[0034] Collect mechanical equipment operation data and perform pre-processing;

[0035] Constructing a detection model based on wavelet scattering forest; wherein the detection model includes a wavelet scattering transform data enhancement module, a deep stacked forest classification module and a similarity measure weighting module;

[0036] Inputting the mechanical equipment data into the detection model, performing data enhancement on the mechanical equipment operation data through the wavelet scattering transform data enhancement module to obtain a plurality of sample sets of different scales;

[0037] Input the sample set into the deep stacked forest classification module for prediction to obtain a number of prediction vectors;

[0038] The similarity measurement weighting module calculates weight vectors at different scales based on the sample set, and performs weighted summation on the prediction vector based on the weight vector to obtain a detection result.

[0039] like Figure 3 The figure shows the detection flow chart within the detection model based on wavelet scattering forest.

[0040] like Figure 4 The figure shows the architecture of the detection model based on wavelet scattering forest.

[0041] In this embodiment, the wavelet scattering transform is used to enhance the operating data of mechanical equipment, generating sample sets at multiple scales. Subsequently, multiple deep stacked forest models are used to train the sample sets at each scale, outputting multiple prediction vectors that are then fused through a weighting mechanism to obtain the final detection result. The weighting mechanism uses cosine similarity to ensure that sample similarity at different scales is fully reflected, thereby optimizing diagnostic effectiveness. The wavelet scattering transform can extract the frequency characteristics of the signal while preserving time domain information, making this method highly tolerant to noise and missing data. Through multi-scale data enhancement, the model can extract data features at different scales, thereby better processing diverse input data and improving the accuracy and reliability of fault diagnosis. In small sample environments, traditional machine learning algorithms are susceptible to sample scarcity. However, by constructing multi-dimensional sample sets, this embodiment can maintain high detection accuracy even in the presence of insufficient data. Furthermore, by extracting fine-grained features from sample sets at different scales and integrating multi-level forest models, this embodiment significantly improves fault diagnosis accuracy and cross-condition adaptability in small sample scenarios.

[0042] In an optional embodiment, the preprocessing step comprises: removing preset abnormal signals in the mechanical equipment operation data; and segmenting and normalizing the mechanical equipment operation data.

[0043] Specifically, according to an actual mechanical equipment fault diagnosis task, sensor is used to acquire diagnosis fault data of the corresponding mechanical equipment, and then a fault data set {X s ,Y s} is obtained through three processes of removing stationary signals, sample segmentation and removing abnormal samples, wherein X S is a vibration signal, Y s is a corresponding label structure, and the label is a fault type corresponding to the sample. The data set is divided into fault training samples for training and fault test data for testing, and then the training samples and the test data are subjected to Z-Score normalization processing.

[0044] In this embodiment, through the abnormal signal rejection and segmentation normalization processing, noise interference and data dimension difference are reduced, high signal-to-noise ratio input is provided for subsequent models, and the reliability of feature extraction is improved.

[0045] In an optional embodiment, the step of performing data enhancement on the mechanical equipment operation data by the wavelet scattering transform data enhancement module to obtain a plurality of sample sets of different scales comprises:

[0046] The mechanical equipment operation data is sequentially subjected to wavelet scattering transform by a plurality of wavelet functions, each wavelet scattering transform result is taken as a single-layer sample set, low-pass filtering is performed on each sample set based on a scale function, and sample sets of different scales are obtained.

[0047] Specifically, for l-layer wavelet scattering transform, a scale function φ and a plurality of wavelet functions are needed when performing transform on the mechanical equipment operation data. λ The wavelet function set is obtained by a series of scaling transformations on the original mother wavelet, and the mother wavelet function ψ(t) in the wavelet scattering transform is:

[0048]

[0049] When performing wavelet scattering transform, the wavelet function is used at the zero layer (the input layer is defined as the zero layer) first, then the wavelet transform is performed on the sample x, and then the obtained wavelet coefficients are taken as a modulus, so that the first layer L1 of the wavelet scattering transform is obtained, and the expression is:

[0050]

[0051] wherein x is the input data.

[0052] Next, the wavelet function is continued to be used for the first layer wavelet coefficient Convolution operation is performed, and the newly obtained wavelet coefficient is taken modulo to obtain the second layer L2 of the wavelet scattering transform, which is expressed as:

[0053]

[0054] By analogy, low-pass filtering is performed on each layer including the zeroth layer using the scaling function φ, and the final output of the wavelet scattering transform, which is called the scattering coefficient, is obtained, and its expression is as follows:

[0055]

[0056] wherein the zeroth layer L0 represents the original sample, and the scattering coefficient S is the deep feature extracted after the wavelet scattering transform of the input data x. The scale used is an adjustable scale parameter. By setting multiple scale parameters, the wavelet scattering transform is performed on the input sample, thereby realizing multi-scale data enhancement.

[0057] In this embodiment, by using different wavelet functions for scattering transform, data features can be extracted in different frequency ranges, thereby increasing the adaptability of the model to complex data patterns. Feature extraction at each scale helps to capture the operation rules of the device under different working conditions, so that the model can more comprehensively and accurately diagnose mechanical faults. Secondly, through low-pass filtering, high-frequency noise can be effectively removed, and important low-frequency features can be retained.

[0058] As an example, the mechanical equipment operation data is transformed three times, and only the second layer scattering coefficient is retained as the result of the data enhancement of the sample. The scattering coefficients are spliced to obtain the data enhanced scattering sample.

[0059] In an optional embodiment, the deep stacked forest classification module comprises a plurality of deep stacked forest models; and the step of inputting the sample set into the deep stacked forest classification module for prediction to obtain a plurality of prediction vectors comprises:

[0060] The sample set of different scales is trained by a plurality of deep stacked forest models, and each deep stacked forest model outputs a prediction vector, and its expression is as follows:

[0061]

[0062] wherein, represents the scattering sample with the Jth scale information, represents the Tth random forest model in the Jth deep stacked forest, and p (1)(X) to p (J) (X) to represent J prediction vectors.

[0063] More specifically, the first layer of random forests in the deep stacked forest is first trained by the sample set, evaluated by out-of-bag data, and each scattered sample outputs a prediction vector representing the probability of the sample belonging to each category. When constructing a deep stacked forest for a certain scale information scattered sample set D, it can be represented as:

[0064] D t = f t t-1 )

[0065] Where f t is the t-th layer of random forest; D1 represents the prediction vector set output by the first layer of random forest to the scattered sample set D through the evaluation of out-of-bag data; D2 represents the prediction vector set output by the second layer to the scattered sample set D, D1, and the prediction vector set output by the t-th layer of random forest. For the scattered sample X S , there are N categories in the target task, and the prediction result F(X) of the deep stacked forest composed of T layers of random forests for the scattered sample X S is:

[0066] F(X) = f T (X S , f1(X S ), f2(X S , f1(X S )), L, f T-1 (X S , L))

[0067] In this embodiment, each deep stacked forest model independently processes data of different scales, and through ensemble learning, the stability of classification can be effectively improved, and complex high-order features in the data can be effectively captured. Even when facing noisy and complex input data, multiple decision trees can be integrated to effectively distinguish and reduce the impact of noise on the final result.

[0068] Further optionally, the step of outputting a prediction vector by the deep stacked forest model comprises: using the decision trees of the random forest to evaluate the performance of each tree through out-of-bag data within the deep stacked forest model, and generating a prediction vector based on the evaluation result.

[0069] In this embodiment, the evaluation method through out-of-bag data can effectively detect the performance of each tree on unseen data, avoiding overfitting of a single tree on the training set. By comprehensively evaluating the performance of different trees, the generalization ability of the final classifier can be improved, making it more stable and accurate in actual application. ​

[0070] In an optional embodiment, the step of calculating the weight vector at different scales based on the sample set by the similarity measure weighting module comprises:

[0071] The sample set average value is calculated at different scales respectively to obtain the class center of each class at different scales;

[0072] The cosine similarity is used as the measure function, the similarity value of the sample set and each class center is calculated at different scales respectively, and the similarity values at the same scale are spliced to obtain the measure vector;

[0073] The measure vector is normalized to obtain the multi-scale weight vector, and the expression is as follows:

[0074]

[0075] Wherein, S represents the cosine similarity function, X S (J) represents the scattering sample with the Jth scale information, and Y i (J) represents the class center of the ith class at the Jth scale, w (1) (X) to w (J) (X) represents J weight vectors.

[0076] As shown in Figure 5 , it is the architecture diagram of the similarity measure weighting module.

[0077] In this embodiment, reasonable distribution of weights helps the model pay more attention to samples closer to the classification boundary, thereby improving the accuracy of fault detection. By splicing and normalizing the similarity values at different scales, the multi-scale information can be effectively fused, further enhancing the comprehensive judgment ability of the model on multiple features, so that the final prediction result is more accurate and robust.

[0078] Further optionally, the step of obtaining the detection result by weighting and summing the prediction vectors based on the weight vector comprises:

[0079] The corresponding prediction vectors are weighted using the multi-scale weight vector, and the average value of the weighted prediction vectors is integrated to obtain the prediction result.

[0080] Specifically, a plurality of weight vectors are used to weight a plurality of prediction vectors, and the expression is as follows:

[0081]

[0082] Wherein, P W(X) is the weighted multiple prediction vectors. Further, since the proposed method adopts the strategy of multi-scale data augmentation and ensemble learning, the weighted multiple prediction vectors need to be integrated. The integration process is realized by calculating the average value of the multiple prediction vectors, which is represented as:

[0083]

[0084] wherein, is the output prediction vector.

[0085] In this embodiment, by weighted summing the prediction results of each scale, the weight of the final result can be adjusted according to the prediction information of different scales, the advantages of each scale are integrated, the accuracy and stability of fault detection are further improved, the prediction deviation under different scales is balanced, the excessive influence of a single scale is reduced, so that the final detection result is more stable and reliable, and more complex and variable actual application scenarios are adapted.

[0086] Embodiment 2

[0087] This embodiment proposes a small sample domain generalization fault detection system based on wavelet scattering forest, which applies the small sample domain generalization fault detection method based on wavelet scattering forest proposed in embodiment 1. As shown in Figure 2 , it is an architecture diagram of a small sample domain generalization fault detection system based on wavelet scattering forest of this embodiment.

[0088] This embodiment proposes a small sample domain generalization fault detection system based on wavelet scattering forest, which applies the small sample domain generalization fault detection method based on wavelet scattering forest proposed in embodiment 1. As shown in

[0089] Data acquisition module: used for collecting mechanical equipment operation data and pre-processing;

[0090] Wavelet scattering transform data augmentation module: used for inputting the mechanical equipment data into the detection model, and performing data augmentation on the mechanical equipment operation data through the wavelet scattering transform data augmentation module to obtain a plurality of sample sets of different scales;

[0091] Deep stacked forest classification module: used for inputting the sample set into the deep stacked forest classification module for prediction to obtain a plurality of prediction vectors;

[0092] Similarity measurement weighting module: used for calculating a weight vector under different scales based on the sample set through the similarity measurement weighting module, and performing weighted sum on the prediction vectors based on the weight vector to obtain a detection result.

[0093] It can be understood that the system of this embodiment corresponds to the method of embodiment 1 described above, and the optional items in embodiment 1 described above are also applicable to this embodiment, so they will not be described here again.

[0094] Example 3

[0095] This example applies the small sample size domain generalization fault detection method based on wavelet scattering forests proposed in Example 1 to public datasets and small datasets for training and prediction.

[0096] In this example, the CNC machine tool spindle bearing dataset (CNC) is used to design the relevant experiment. In this experimental platform, the faulty bearings are installed on the spindle. Secondly, in order to collect data, an accelerometer is installed on the machine tool spindle to collect vibration signal data of the spindle bearing at a sampling frequency of 25 kHz under normal service conditions of the machine tool, which covers three working conditions of cutting aluminum materials at 6000, 7000 and 8000 rpm of the machine tool respectively. The spindle bearings used in the CNC dataset are angular contact ball bearings, the model of which is NSK 40BNR10. The bearing faults are artificially implanted by means of electric spark machining and laser marking, and there are differences in location between the faults, including inner ring (IR) fault, outer ring (OR) fault and cage ring (CR) fault.

[0097] To construct the task in the working condition generalization scenario, four kinds of bearing data in the CNC dataset are selected to construct the small sample fault diagnosis task, covering four bearing states including healthy, inner ring (IR) fault, outer ring (OR) fault and cage ring (CR) fault, which are represented by class 1 to class 4 respectively. In the experiment, non-overlapping sliding windows are used to collect samples on the time domain vibration signals of the bearings provided by the dataset, and 450 data points of 2048 length samples are collected for each bearing (150 for each working condition). Then, small sample fault diagnosis tasks are constructed under two working condition generalization scenarios, simulating the working condition changes of the machine tool spindle speed from 7000 rpm to 8000 rpm (scenario 1) and from 8000 rpm to 7000 rpm (scenario 2). Among them, scenario 1 means that each method needs to use a small number of labeled samples in the 7000 rpm working condition in the dataset to train the model, and extend the diagnosis ability of the model to the target working condition of 8000 rpm, to realize effective fault diagnosis for the samples in this working condition. The wavelet scattering forest network parameters constructed are shown in Table 1:

[0098] Table 1 Hyperparameter settings of the detection model based on wavelet scattering forests

[0099]

[0100]

[0101] The random forest hyperparameters include 500 decision trees, the maximum feature selection number is sqrt, the minimum leaf node sample number is 2, and the node splitting criterion is gini. In the integrated model part, the scale number is 5, the scale parameters are 3, 4, 5, 6 and 7, the number of wavelets per octave is 12, and the number of random forest stacking layers is 4. In addition, in the training stage, the model uses the cross-entropy loss function and uses the Adam optimizer for optimization. The learning rate is initially set to 0.001, the batch size is 50, and the total iteration is 300. To improve the convergence and prevent overfitting, a learning rate decay strategy is introduced, and the decay factor is set to 0.8.

[0102] The method of the application is compared with other fault diagnosis methods:

[0103] To verify the superiority of the proposed method, the following five fault diagnosis methods are used for comparison, including: random forest (RF), deep forest (DF), wide kernel deep convolutional neural network (WDCNN), wide kernel deep convolutional neural network + fine-tuning (WDCNN-FT) and prototype network (PN). For the methods based on transfer learning and meta-learning, 6000 rpm data is additionally provided as the source domain to assist model training. According to the above settings, the data used by each method can be divided into three domains: source domain, intermediate domain and target domain. Among them, the source domain has rich labeled samples, the intermediate domain only has a small amount of labeled samples, and the target domain has no labeled samples, and finally the target domain needs to achieve accurate fault diagnosis, and finally the sample number used by each method in 1-shot, 2-shot and 5-shot tasks is shown in Table 2. As Figure 6 shown, it is the precision comparison chart of the method of the application and other fault diagnosis methods.

[0104] To reduce the randomness of the experiment, 10 repeated experiments are performed for each task, and the average value is taken as the final experimental result. According to the above settings, the 4-way K-shot small sample fault diagnosis experiment is carried out in the generalization scene of two working conditions using the CNC data set. The test accuracy of each method is shown in Table 3 and Figure 6The experimental results show that in the small sample fault diagnosis task of working condition generalization, the highest test accuracy of the method of the application is more than 98%, which is much better than other comparative methods, showing that it can guarantee high diagnostic accuracy and excellent generalization ability under very few training data. Although WDCNN-FT based on transfer learning and PN based on meta-learning also obtained higher accuracy in part of the task (97.10% and 94.23% respectively in the 1-shot task of scene 2), they rely on additional source domain data, and the practical application convenience is not as good as the method of the application. On the contrary, traditional methods such as RF and WDCNN have lower test accuracy because they do not introduce source domain data and lack data enhancement strategy. In the 1-shot task of scene 1 and scene 2, the accuracy of RF is only 26.32% and 27.03%; WDCNN only achieves an accuracy of 41.02% in the 1-shot task of scene 1, which is significantly lower than the improved WDCNN-FT after transfer learning. In addition, although the DF method also uses multi-scale data enhancement and a deep stacking model of random forests, its test accuracy has always been less than 70%. This is mainly because with the increase of the number of layers, the shallow information is gradually "diluted", making it difficult for the deep model to obtain sufficient data information. The method of the application, however, integrates the outputs of all previous layers in the deep random forest, achieving full information transmission, thus better exploiting the advantages of deep models and exhibiting excellent diagnostic and generalization performance under small sample conditions.

[0105] Table 2 Number of samples used by each method

[0106]

[0107] Table 3 Quantitative comparison of average diagnostic accuracy (%) of different models

[0108]

[0109]

[0110] In the foregoing experiments, the effectiveness and advancement of the proposed model have been preliminarily verified. Further, considering the characteristics that rotating machinery in real industrial environments will also be affected by different levels of environmental noise during operation, small sample fault diagnosis experiments under working condition generalization and strong noise scenarios are designed to explore whether the proposed model has sufficient superior generalization performance. This experiment is based on the 5-shot diagnosis task under two working condition generalization scenarios, and Gaussian white noise is added to the training and test samples of each method in this task, with a signal-to-noise ratio of 0 dB, 2 dB and 4 dB. In addition, the other settings in this experiment are consistent with the settings of the foregoing experiments. Under the working condition generalization and strong noise scenario, the test accuracy of each method is shown in Table 4.

[0111] From the experimental results of Table 4 and Figure 6 It can be seen that, under the interference of noise, the accuracy of RF and WDCNN is at the lowest level, which is difficult to meet the demand of small sample variable working condition fault diagnosis; and the PN has the most significant reduction, such as the accuracy of scene 1 at 0dB is reduced by 33.74% compared with no noise, which may be due to the change of sample feature distribution caused by noise, resulting in the deviation of measurement results. In contrast, WDCNN-FT relies on pre-training fine-tuning, and DF through multi-scale integrated learning, both show strong anti-noise ability. The method of the present application still maintains the highest accuracy in most tasks, only slightly inferior to WDCNN-FT at 4dB in scene 2, and does not need to rely on source domain data, which is more convenient to use, has obvious advantages and competitiveness.

[0112] Table 4 Quantitative comparison of average diagnostic accuracy (%) of different models after adding noise

[0113]

[0114]

[0115] Example 4

[0116] In this embodiment, the small sample domain generalization fault detection method based on wavelet scattering forest proposed in Example 1 is applied to the ablation experiment of public data set and small data set.

[0117] In this embodiment, an ablation experiment is designed using the CNC data set. Five ablation models based on the wavelet scattering forest detection model are designed, represented by AM1 to AM5. Among them, AM1 only contains the classification module of the wavelet scattering forest detection model, while AM2 contains both the classification module and the weighting module. AM3 is composed of a single-scale data enhancement module and a classification module, and AM4 further combines a weighting module based on AM3. AM5 is composed of a multi-scale data enhancement module (5 scales are set in the experiment) and a classification module, but does not contain a weighting module.

[0118] As shown in Figure 7 , it is the precision comparison chart of the ablation experiment of the method of the present application.

[0119] Figure 7The test accuracy of each ablation model and the complete wavelet scattering forest-based detection model in the 1-shot task under scenario 1 is shown. The experimental results show that the accuracy of AM1 is only 27.69%, indicating that the single classification ability is difficult to meet the small sample fault diagnosis demand. After introducing the weighted module (AM2) and the data enhancement module (AM3), the accuracy is increased by 42.16% and 67.44% respectively than AM1, which verifies the significant effect of the two modules in improving the diagnostic accuracy. Further, the use of multi-scale data enhancement (AM5) compared with single scale (AM3) increases the accuracy by 3.18%, which benefits from the wavelet scattering transform that can extract more multi-scale features, thereby training multiple deep stacked random forests with diversity. In addition, AM4 and the final wavelet scattering forest-based detection model are increased by 2.82% and 0.65% respectively than AM3 and AM5, proving that the synergistic effect of data enhancement and weighting module is more effective. In summary, the integration of classification, weighting and multi-scale data enhancement modules significantly improves the performance and accuracy of the model in small sample variable working condition fault diagnosis.

[0120] Embodiment 5

[0121] The embodiment provides a computer device, including a memory and a processor, and the memory stores computer readable instructions, wherein the computer readable instructions are executed by the processor to make the processor execute the steps of the small sample domain generalization fault detection method based on the wavelet scattering forest provided in the embodiment 1.

[0122] Embodiment 6

[0123] The embodiment provides a storage medium, which stores computer readable instructions, wherein the computer readable instructions are executed by a processor to realize the steps of the small sample domain generalization fault detection method based on the wavelet scattering forest provided in the embodiment 1.

[0124] Exemplarily, the storage medium includes but is not limited to a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk and various media that can store program codes.

[0125] Exemplarily, the instructions, programs, code sets or instruction sets can be implemented by using a conventional programming language.

[0126] Exemplarily, the processor includes but is not limited to a smart phone, a personal computer, a server, a network device and the like, and is used to execute all or part of the steps of the small sample domain generalization fault detection method based on the wavelet scattering forest described in the embodiment 1.

Claims

1. A small sample size domain generalization fault detection method based on wavelet scattering forests, characterized in that, The method comprises the following steps: Collecting mechanical equipment operation data and preprocessing; Constructing a detection model based on wavelet scattering forest; wherein the detection model comprises a wavelet scattering transformation data enhancement module, a deep stacked forest classification module, and a similarity measurement weighting module; Inputting the mechanical equipment data into the detection model, performing data enhancement on the mechanical equipment operation data through the wavelet scattering transformation data enhancement module, and obtaining a plurality of sample sets of different scales; Inputting the sample sets into the deep stacked forest classification module for prediction to obtain a plurality of prediction vectors; Calculating a weight vector under different scales based on the sample sets through the similarity measurement weighting module, and performing weighted summation on the prediction vectors based on the weight vector to obtain a detection result.

2. The small sample size domain generalization fault detection method based on wavelet scattering forests according to claim 1, characterized in that, The preprocessing step comprises: removing preset abnormal signals in the mechanical equipment operation data; and performing normalization processing on the segmented mechanical equipment operation data.

3. The method according to claim 1, wherein, The step of performing data enhancement on the mechanical equipment operation data through the wavelet scattering transformation data enhancement module to obtain a plurality of sample sets of different scales comprises: Performing wavelet scattering transformation on the mechanical equipment operation data through a plurality of wavelet functions in sequence, taking each wavelet scattering transformation result as a single-layer sample set, performing low-pass filtering on any sample set based on a scale function, and obtaining sample sets of different scales.

4. The small sample size domain generalization fault detection method based on wavelet scattering forests according to claim 1, characterized in that, The deep stacked forest classification module comprises a plurality of deep stacked forest models; and the step of inputting the sample sets into the deep stacked forest classification module for prediction to obtain a plurality of prediction vectors comprises: Performing prediction on sample sets of different scales through a plurality of deep stacked forest models respectively, and outputting one prediction vector by each deep stacked forest model, which has the following expression: wherein, represents a scattering sample with the J-th scale information, represents the T-th layer random forest model inside the J-th deep stacked forest, p (1) (X) to p (J) (X) to represent J prediction vectors.

5. The small sample size domain generalization fault detection method based on wavelet scattering forests according to claim 4, characterized in that, The step of outputting one prediction vector by the deep stacked forest model comprises: using a decision tree of a random forest to evaluate the performance of each tree through out-of-bag data in the deep stacked forest model, and generating a prediction vector based on the evaluation result.

6. The small sample size domain generalization fault detection method based on wavelet scattering forests according to any one of claims 1-5, characterized in that, The step of calculating a weight vector under different scales based on the sample sets through the similarity measurement weighting module comprises: Calculating the average value of the sample sets on different scales to obtain the class center of each class on different scales; Using cosine similarity as a measurement function to calculate the similarity values of the sample sets and each class center on different scales, and splicing the similarity values of the same scale to obtain a measurement vector; Normalizing the measurement vector to obtain a multi-scale weight vector, which has the following expression: where S represents a cosine similarity function, represents a scattering sample with the J-th scale information, and Y i (J) represents a class center of the i-th class on the J-th scale, w (1) (X) to w (J) (X) represents J weight vectors.

7. The method according to claim 6, wherein, The step of performing weighted summation on the prediction vectors based on the weight vector to obtain a detection result comprises: Performing weighting on the corresponding prediction vectors by using the multi-scale weight vector, and integrating the average values of the weighted prediction vectors to obtain a prediction result.

8. A small sample size domain generalization fault detection system based on wavelet scattering forests, applying the method of any one of claims 1 to 7, characterized in that, Comprise: A data collection module for collecting mechanical equipment operation data and preprocessing; A wavelet scattering transformation data enhancement module for inputting the mechanical equipment data into the detection model, performing data enhancement on the mechanical equipment operation data through the wavelet scattering transformation data enhancement module, and obtaining a plurality of sample sets of different scales; a deep stacked forest classification module, configured to input the sample set into the deep stacked forest classification module for prediction to obtain a plurality of prediction vectors; a similarity measure weighting module, configured to calculate a weight vector at different scales based on the sample set through the similarity measure weighting module, and perform weighted summation on the prediction vectors based on the weight vector to obtain a detection result.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the small sample domain generalization fault detection method based on the wavelet scattering forest in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the small sample domain generalization fault detection method based on the wavelet scattering forest in any one of claims 1-7.

Images