Cross-device fault diagnosis method based on heterogeneous signals

Abstract

The application discloses a cross-device fault diagnosis method based on heterogeneous signals and belongs to the technical field of cross-device fault diagnosis. The method comprises the following steps: constructing an acoustic characteristic data set and a multi-source vibration characteristic data set according to acoustic signals and vibration signals in a plurality of source domain vibration data sets; dividing the multi-source vibration characteristic data set into a first training set and a first test set according to a first preset proportion at random; constructing a small sample acoustic vibration pairing training set, a second verification set and a second test set according to the acoustic characteristic data set and the multi-source vibration characteristic data set; pretraining and testing a vibration teacher model according to the first training set and the first test set to generate a pretraining vibration teacher model; and guiding a lightweight student model to train on acoustic signals by using vibration knowledge extracted by the pretraining vibration teacher model according to the small sample acoustic vibration pairing training set and the second verification set. The application realizes abnormal detection based on acoustic signals under the conditions of small samples and cross devices.

Description

Technical Field

[0001] This invention relates to a cross-device fault diagnosis method based on heterogeneous signals, belonging to the field of cross-device fault diagnosis technology. Background Technology

[0002] With the continuous improvement of the intelligence level of industrial equipment, real-time status monitoring and accurate fault diagnosis of key rotating components such as rolling bearings have become crucial for ensuring production safety and avoiding significant economic losses. Among various monitoring methods, vibration signal-based analysis and diagnosis technology has been widely used in industry and has become the mainstream technology for fault diagnosis due to its ability to provide rich information on the internal status of equipment, high feature recognition, and relatively strong resistance to environmental interference. However, this technology has significant limitations in terms of physical deployment. For many critical equipment operating in complex or harsh environments, such as the rolling mill bearings in the non-ferrous metallurgical industry that are subjected to high temperatures, high humidity, and limited space, the installation of contact vibration sensors often faces great difficulties or is even completely infeasible. This physical limitation directly leads to the inability to effectively obtain key status information of equipment operation, causing vibration signal-based diagnostic methods to fail in such scenarios. As an important supplement to non-contact monitoring, acoustic signals collected by microphones, with their flexible deployment advantages, provide a potential monitoring approach for the aforementioned equipment where vibration sensors cannot be installed. However, acoustic signals face severe challenges in industrial settings, especially in high-noise and high-interference environments such as rolling mills in the non-ferrous metallurgical industry: significant environmental noise can easily drown out weak acoustic features related to faults, making feature extraction difficult; at the same time, the propagation path of sound waves is extremely complex due to factors such as equipment structure, cavity and medium, which makes the feature distribution of acoustic signals vary greatly even under the same equipment or different operating conditions, seriously weakening the generalization ability of diagnostic models in actual cross-equipment applications.

[0003] Existing research on cross-device fault diagnosis primarily focuses on scenarios where the source and target devices use the same type of signal, particularly in vibration-to-vibration signal transfer scenarios, attempting to bridge the feature distribution differences between devices through domain adaptation algorithms. However, when the source device possesses abundant vibration signal data while the target device can only acquire acoustic signals, the fundamental modal differences between vibration and acoustic signals in terms of physical mechanisms and feature representations render these isomorphic signal transfer-based methods ineffective, resulting in knowledge transfer from the source domain actually harming the performance of the target domain. Furthermore, in industrial settings, fault samples for target devices, especially new or critical equipment, are typically extremely scarce, while the massive vibration data accumulated by the source device cannot be directly used to train acoustic diagnostic models for the target domain due to modal differences. This data imbalance problem of scarce target domain fault samples is further amplified in cross-modal scenarios, leading to models trained directly on limited and heavily noise-contaminated acoustic samples being prone to overfitting and exhibiting severely inadequate diagnostic generalization performance. Therefore, it is urgent to break through the limitations of existing methods and develop new cross-device fault diagnosis methods that can effectively utilize vibration signal knowledge, overcome the inherent defects of acoustic signals, and adapt to the small sample conditions of target equipment.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] This invention provides a cross-device fault diagnosis method based on heterogeneous signals. The method deploys easily installed non-contact acoustic sensors on the target processing equipment to construct a small-sample acoustic training set, and uses vibration signal data from multiple sources as a source of transfer knowledge to guide the analysis and diagnosis of acoustic signals of the target equipment. Ultimately, it achieves anomaly detection based on acoustic signals under small-sample and cross-device conditions.

[0006] The technical solution of this invention is:

[0007] A cross-device fault diagnosis method based on heterogeneous signals includes:

[0008] S1. Use contact methods to collect vibration signals of the test object on multiple source devices and construct multiple source domain vibration datasets; use non-contact methods to collect acoustic signals of the test object on the target device.

[0009] S2. Based on the acoustic signal and the vibration signals in the multiple source domain vibration datasets, construct an acoustic feature dataset and a multi-source vibration feature dataset;

[0010] S3. The multi-source vibration feature dataset is randomly divided into a first training set and a first test set according to a first preset ratio;

[0011] S4. Based on the acoustic feature dataset and the multi-source vibration feature dataset, construct a small-sample acoustic-vibration paired training set, a second validation set, and a second test set;

[0012] S5. Construct a knowledge distillation framework, which includes a vibration teacher model and a lightweight student model; pre-train and test the vibration teacher model based on the first training set and the first test set to generate a pre-trained vibration teacher model; based on a small sample acoustic-vibration paired training set and a second validation set, use the vibration knowledge extracted from the pre-trained vibration teacher model to guide the lightweight student model to train on acoustic signals, thereby realizing vibration knowledge transfer.

[0013] Further, S1 includes:

[0014] S11. Select a device that is compatible with the operating conditions of the target device and has the same type of test object as the source device; and deploy contact acceleration sensors on the test objects of multiple source devices.

[0015] S12. Vibration signals of multiple source devices are collected from the test object using a contact accelerometer; acoustic signals of the target device are collected from the test object using a non-contact microphone.

[0016] S13. Based on the operating condition parameters of the target equipment, select a subset of data that matches the operating conditions from the vibration signals of multiple source equipment, and construct multiple source domain vibration datasets that are aligned with the operating conditions.

[0017] Furthermore, the adaptation to the target equipment's operating conditions specifically includes:

[0018] Using rotational speed as the target equipment's operating condition parameter, select the equipment whose operating condition parameter value differs from the target equipment's value within a preset threshold as the source equipment.

[0019] Furthermore, S2 specifically includes:

[0020] S21. The acoustic signal and the vibration signals in the multiple source domain vibration datasets are respectively augmented using the sliding window method to generate an augmented time-domain acoustic signal sample set and multiple source time-domain vibration signal sample sets.

[0021] S22. Within a preset frequency band, extract 1 / 3 octave band features on the corresponding center frequency point bandwidth from the samples in the enhanced time-domain acoustic signal sample set and the samples in the multiple source time-domain vibration signal sample sets.

[0022] S23. Perform feature dimension truncation processing on the 1 / 3 octave band feature dimension of each of the multiple source devices and the 1 / 3 octave band feature dimension of the target device to obtain the truncated 1 / 3 octave band features of the multiple source devices and the truncated 1 / 3 octave band features of the target device.

[0023] S24. Preprocess the truncated 1 / 3 octave band features of the multiple source devices and the truncated 1 / 3 octave band features of the target device to obtain preprocessed 1 / 3 octave band features; integrate the preprocessed 1 / 3 octave band features of the multiple source devices to construct a multi-source vibration feature dataset, and construct an acoustic feature dataset from the preprocessed 1 / 3 octave band features of the target device.

[0024] Furthermore, S23 specifically includes:

[0025] The minimum value of the 1 / 3 octave band characteristic dimension among the multiple source devices is taken as the first dimension value; the source device with the 1 / 3 octave band characteristic dimension value as the first dimension value is taken as the first source device, and the rest are second source devices; the 1 / 3 octave band characteristic dimension value of the target device is taken as the second dimension value.

[0026] The first dimension value and the second dimension value are compared: if the first dimension value is less than the second dimension value, the 1 / 3 octave band feature dimension of the second source device and the target device is truncated so that the truncated 1 / 3 octave band feature dimension value is the first dimension value; otherwise, the 1 / 3 octave band feature dimension of the second source device is truncated so that the truncated 1 / 3 octave band feature dimension value is the first dimension value.

[0027] Further, S4 includes:

[0028] S41. Construction of a small-sample acoustic training set: From the acoustic feature dataset, M normal category acoustic feature samples with a first preset proportion and N fault category acoustic feature samples with a second preset proportion are extracted according to the floor rule to form a small-sample acoustic training set; wherein, the first preset proportion is greater than the second preset proportion.

[0029] S42. Construction of small sample acoustic-vibration pairing training set: Based on the small sample acoustic training set, randomly select M normal category labels and N fault category labels from the multi-source vibration feature dataset, and pair them one by one with the acoustic samples with the same label in the small sample acoustic training set to construct the small sample acoustic-vibration pairing training set.

[0030] S43. The samples in the acoustic feature dataset, excluding the small sample acoustic training set, are randomly divided into a second validation set and a second test set according to the sample category labels and a second preset ratio.

[0031] Furthermore, S5 specifically includes:

[0032] S51. Construct a knowledge distillation framework, which includes a vibrating teacher model and a lightweight student model. The vibrating teacher model uses ResNet-18 as its basic architecture and performs a dimensionality reduction transformation on the convolutional layers in ResNet-18 from two dimensions to one dimension: replacing the two-dimensional convolutional kernels with one-dimensional convolutional kernels and reducing the channel dimension of the convolutional layers. Through this transformation, a lightweight one-dimensional improved architecture 1D-ResNet-18 is constructed as the vibrating teacher model. The lightweight student model is based on the lightweight one-dimensional improved architecture 1D-ResNet-18. By reducing the number of residual blocks to one and introducing a max pooling layer after the residual blocks, a lightweight student model 1D-ResNet-6 is constructed.

[0033] S52. Vibration teacher model pre-training: Based on the first training set and the first test set, the vibration teacher model is pre-trained and tested to generate a pre-trained vibration teacher model.

[0034] S53. During knowledge distillation training on a small sample acoustic-vibration pairing training set, in each epoch of training, the pre-trained vibration teacher model receives vibration samples from the small sample acoustic-vibration pairing training set as input, while the lightweight student model receives acoustic samples with the same category label from the small sample acoustic-vibration pairing training set as input. The distillation loss is formed by the temperature-softened soft label output by the pre-trained vibration teacher model and the temperature-softened soft label predicted by the lightweight student model. The hard loss is calculated by comparing the label predicted by the lightweight student model with the true label. Finally, the total loss function including the distillation loss and the hard loss is obtained. Then, the model is validated on the second validation set, and the hyperparameters in the knowledge distillation process are optimized to obtain the enhanced student model after knowledge distillation transfer.

[0035] The beneficial effects of this invention are:

[0036] I. This invention avoids the physical dependence on contact vibration sensors by adopting a convenient non-contact acoustic signal acquisition scheme. This design shows significant advantages for sensor deployment in many critical equipment operating in complex or harsh environments, such as the high-temperature, high-humidity, and space-constrained roll bearings in rolling mills in the non-ferrous metallurgical industry, providing a new technical path for anomaly detection in critical equipment in manufacturing.

[0037] Second, this invention utilizes the clear prior knowledge of faults contained in the vibration signals of multiple source devices with operating conditions adapted to the target device and of the same test object type. Through a cross-modal knowledge transfer mechanism, it guides the analysis and diagnosis of the target device's acoustic signals. This strategy significantly suppresses the susceptibility of acoustic signals to interference in industrial noise environments, fundamentally enhancing the robustness of the acoustic diagnostic model and overcoming the bottleneck of insufficient generalization ability of traditional single acoustic monitoring models in small sample scenarios.

[0038] Third, this invention significantly bridges the modal gap of heterogeneous acoustic and vibration signals from the perspective of frequency domain energy distribution by extracting 1 / 3 octave band features. Furthermore, by combining feature dimension truncation processing with a small-sample acoustic training set construction strategy, it improves cross-device feature space compatibility while specifically addressing the severe challenge of scarce fault samples in industrial settings, laying a data foundation for reliable diagnosis under small-sample conditions.

[0039] Fourth, this invention utilizes knowledge distillation technology to simultaneously achieve cross-modal migration from vibration signals to acoustic signals and cross-device migration from multiple source devices to the target device. This collaborative migration mechanism significantly reduces the negative migration risk caused by direct transmission of heterogeneous signals and effectively suppresses the overfitting tendency of acoustic diagnostic models under conditions of scarce fault samples in the vibration signals of multiple source devices by leveraging the rich prior fault knowledge in the vibration signals of multiple source devices, thus significantly improving the diagnostic generalization performance in small sample scenarios. Furthermore, this invention achieves efficient real-time monitoring in industrial scenarios through a lightweight model architecture and octave band feature optimization. Experimental verification shows that this method, while ensuring diagnostic accuracy, has a short model inference time and can be deployed on edge computing devices, meeting the needs of parallel monitoring of multiple devices in complex industrial scenarios. Compared to traditional methods that rely on complex sensor networks or high-cost data integration, this invention significantly lowers the implementation threshold.

[0040] In summary, this invention circumvents sensor deployment limitations through a non-contact acoustic acquisition scheme, enhances the robustness of acoustic diagnosis by utilizing a multi-source vibration signal cross-modal transfer mechanism, and overcomes the dual bottlenecks of modal gap and sample scarcity based on 1 / 3 octave band feature fusion and knowledge distillation techniques. This method significantly reduces the risk of negative transfer while achieving microsecond-level inference efficiency per sample through lightweight model design, solving problems such as poor generalization of acoustic signal diagnosis in noisy environments and model overfitting due to the scarcity of fault samples in target equipment. Its edge-deployable nature provides a modification-free real-time monitoring solution for harsh operating conditions such as rolling mills in the non-ferrous metallurgical industry, significantly reducing implementation costs while ensuring diagnostic accuracy, and has significant engineering value for improving the status awareness capabilities of key equipment in the manufacturing industry. Attached Figure Description

[0041] Figure 1 This is a flowchart of the cross-device fault diagnosis method based on heterogeneous signals according to the present invention.

[0042] Figure 2 This is a schematic diagram of the accuracy and F1 score results of the second validation set in an embodiment of the present invention.

[0043] Figure 3 This is a schematic diagram of the confusion matrix results of the second test set in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0045] Example 1: As Figures 1-3 As shown, according to a first aspect of the present invention, a cross-device fault diagnosis method based on heterogeneous signals is provided, comprising:

[0046] S1. Use contact methods to collect vibration signals of the test object on multiple source devices and construct multiple source domain vibration datasets; use non-contact methods to collect acoustic signals of the test object on the target device.

[0047] Furthermore, S1 specifically includes:

[0048] S11. Select a device that is compatible with the operating conditions of the target device and has the same type of test object as the source device; and deploy contact acceleration sensors on the test objects of multiple source devices.

[0049] S12. Vibration signals of multiple source devices are collected from the test object using a contact accelerometer; acoustic signals of the target device are collected from the test object using a non-contact microphone.

[0050] S13. Based on the operating condition parameters of the target equipment, select a subset of data that matches the operating conditions from the vibration signals of multiple source equipment, and construct multiple source domain vibration datasets that are aligned with the operating conditions.

[0051] Furthermore, the adaptation to the target equipment's operating conditions specifically includes:

[0052] Using rotational speed as the target equipment's operating condition parameter, select the equipment whose operating condition parameter value differs from the target equipment's value within a preset threshold as the source equipment.

[0053] For example, the target device is a rotating device, which can be a bearing. If the target device's rotational speed is 1750 rpm, and the preset threshold for rotational speed is 300 rpm, then the selected source device's rotational speed range is 1450 rpm to 2050 rpm. If the target device's test object type is a deep groove ball bearing, then the source device's test object type is also a deep groove ball bearing.

[0054] S2. Based on the acoustic signal and the vibration signals in the multiple source domain vibration datasets, construct an acoustic feature dataset and a multi-source vibration feature dataset;

[0055] Furthermore, S2 specifically includes:

[0056] S21. For the acoustic signal and the vibration signals in the multiple source domain vibration datasets, data augmentation is performed using the sliding window method to generate an augmented time-domain acoustic signal sample set and multiple source time-domain vibration signal sample sets; the formula for calculating the size of the data augmentation sample set is as follows:

[0057] ;

[0058] Where N is the size of the augmented sample set; L is the length of the original signal; W is the window length; and S is the step size, S=0.1 L.

[0059] S22. Within a preset frequency band, extract the 1 / 3 octave band features on the bandwidth of the corresponding center frequency point from the samples in the enhanced time-domain acoustic signal sample set and the multiple source time-domain vibration signal sample sets; the 1 / 3 octave band features are obtained from the root mean square value of the 1 / 3 octave band power spectrum, assuming the time-domain vibration / acoustic signal... The autocorrelation function is bilateral self-power spectrum The power spectral density function is ,but:

[0060] ;

[0061] ;

[0062] Further calculations can be performed using 1 / 3 octave band characteristics. :

[0063] ;

[0064] in, Indicates frequency; This is the i-th 1 / 3 octave band feature; The lower limit frequency of 1 / 3 octave band; It is the upper limit frequency of 1 / 3 octave band.

[0065] S23. Perform feature dimension truncation processing on the 1 / 3 octave band feature dimension of each of the multiple source devices and the 1 / 3 octave band feature dimension of the target device to obtain the truncated 1 / 3 octave band features of the multiple source devices and the truncated 1 / 3 octave band features of the target device.

[0066] Furthermore, S23 specifically includes:

[0067] The minimum value of the 1 / 3 octave band characteristic dimension among the multiple source devices is taken as the first dimension value; the source device with the 1 / 3 octave band characteristic dimension value as the first dimension value is taken as the first source device, and the rest are second source devices; the 1 / 3 octave band characteristic dimension value of the target device is taken as the second dimension value.

[0068] The first dimension value and the second dimension value are compared: if the first dimension value is less than the second dimension value, the 1 / 3 octave band feature dimension of the second source device and the target device is truncated so that the truncated 1 / 3 octave band feature dimension value is the first dimension value; otherwise, the 1 / 3 octave band feature dimension of the second source device is truncated so that the truncated 1 / 3 octave band feature dimension value is the first dimension value, and the 1 / 3 octave band feature dimension value of the target device remains unchanged.

[0069] S24. Preprocess the truncated 1 / 3 octave band features of the multiple source devices and the truncated 1 / 3 octave band features of the target device to obtain preprocessed 1 / 3 octave band features; integrate the preprocessed 1 / 3 octave band features of the multiple source devices to construct a multi-source vibration feature dataset, and construct an acoustic feature dataset from the preprocessed 1 / 3 octave band features of the target device.

[0070] S3. The multi-source vibration feature dataset is randomly divided into a first training set and a first test set according to a first preset ratio;

[0071] S4. Based on the acoustic feature dataset and the multi-source vibration feature dataset, construct a small-sample acoustic-vibration paired training set, a second validation set, and a second test set;

[0072] Furthermore, S4 specifically includes:

[0073] S41. Construction of a small-sample acoustic training set: From the acoustic feature dataset, M normal category acoustic feature samples with a first preset proportion and N fault category acoustic feature samples with a second preset proportion are extracted according to the floor rule to form a small-sample acoustic training set; wherein, the first preset proportion is greater than the second preset proportion.

[0074] S42. Construction of small sample acoustic-vibration pairing training set: Based on the small sample acoustic training set, randomly select M normal category labels and N fault category labels from the multi-source vibration feature dataset, and pair them one by one with the acoustic samples with the same label in the small sample acoustic training set to construct the small sample acoustic-vibration pairing training set.

[0075] S43. The samples in the acoustic feature dataset, excluding the small sample acoustic training set, are randomly divided into a second validation set and a second test set according to the sample category labels and a second preset ratio.

[0076] S5. Construct a knowledge distillation framework, which includes a vibration teacher model and a lightweight student model; pre-train and test the vibration teacher model based on the first training set and the first test set to generate a pre-trained vibration teacher model; based on a small sample acoustic-vibration paired training set and a second validation set, use the vibration knowledge extracted from the pre-trained vibration teacher model to guide the lightweight student model to train on acoustic signals, thereby realizing vibration knowledge transfer.

[0077] Furthermore, S5 specifically includes:

[0078] S51. Constructing a knowledge distillation framework: The vibration teacher model uses ResNet-18 as its basic architecture, leveraging its residual connection structure to enhance deep feature extraction capabilities; dimensionality reduction is implemented in the convolutional layers of ResNet-18 from two dimensions to one dimension: two-dimensional convolutional kernels are replaced with one-dimensional convolutional kernels, and the channel dimension of the convolutional layers is reduced to control the number of parameters, adapting to the one-dimensional characteristics of 1 / 3 octave band feature vectors and the limited feature dimension; a lightweight one-dimensional improved architecture 1D-ResNet-18 is constructed as the vibration teacher model through this modification; the lightweight student model is based on the lightweight one-dimensional improved architecture 1D-ResNet-18, and by reducing the number of residual blocks to one and introducing a max pooling layer after the residual blocks, a lightweight student model 1D-ResNet-6 is constructed;

[0079] S52. Vibration teacher model pre-training: Based on the first training set and the first test set, the vibration teacher model is pre-trained and tested to learn the fault characterization knowledge of vibration signals and generate a pre-trained vibration teacher model.

[0080] S53. During knowledge distillation training on a small sample acoustic-vibration pairing training set, in each epoch, the pre-trained vibration teacher model receives vibration samples from the small sample acoustic-vibration pairing training set as input, while the lightweight student model receives acoustic samples with the same category label from the small sample acoustic-vibration pairing training set as input. The distillation loss is constituted by the temperature-softened soft label output by the pre-trained vibration teacher model and the temperature-softened soft label predicted by the lightweight student model. The hard loss is calculated by comparing the label predicted by the lightweight student model with the true label. Finally, the total loss function including distillation loss and hard loss is obtained. Then, the knowledge distillation is validated on the second validation set, and the hyperparameters (including temperature T, loss ratio, etc.) in the knowledge distillation process are optimized. This process ultimately yields an enhanced student model that has undergone knowledge distillation and transfer. The enhanced student model is then tested using a second test set for performance evaluation. Furthermore, the enhanced student model is used to diagnose the acoustic signals acquired on the target device.

[0081] Furthermore, it also includes S54, independent training of lightweight student models: based on the acoustic samples in the small sample acoustic vibration pairing training set (i.e., the small sample acoustic training set), the lightweight student model is trained independently to obtain the baseline student model without knowledge transfer, which is used for the comparison and verification of the subsequent knowledge distillation effect.

[0082] The student model label softening is adjusted using a temperature parameter T, which modifies the logical value of each category. Convert to normalized probability value :

[0083] ;

[0084] The total loss function is shown below:

[0085] ;

[0086] in, For soft label weighting coefficients; T is the temperature parameter; For the teacher model, the soft label probability; This is a real label. In this embodiment... =0.5; T=13.

[0087] According to a second aspect of the present invention, a cross-device fault diagnosis system based on heterogeneous signals is provided, comprising: a first module configured to perform S1: acquiring vibration signals of a test object on multiple source devices using a contact method to construct multiple source domain vibration datasets; acquiring acoustic signals of a test object on a target device using a non-contact method; a second module configured to perform S2: constructing an acoustic feature dataset and a multi-source vibration feature dataset based on the acoustic signals and vibration signals in the multiple source domain vibration datasets; and a third module configured to perform S3: randomly dividing the multi-source vibration feature dataset into a first training set and a... The first test set; the fourth module, used to execute S4: constructing a small-sample acoustic-vibration pairing training set, a second validation set, and a second test set based on the acoustic feature dataset and the multi-source vibration feature dataset; the fifth module, used to execute S5: constructing a knowledge distillation framework, which includes a vibration teacher model and a lightweight student model; pre-training and testing the vibration teacher model based on the first training set and the first test set to generate a pre-trained vibration teacher model; using the vibration knowledge extracted from the pre-trained vibration teacher model based on the small-sample acoustic-vibration pairing training set and the second validation set to guide the lightweight student model in training on acoustic signals, thereby achieving vibration knowledge transfer. For parts of each module not described in detail above, please refer to the relevant descriptions in the embodiments.

[0088] According to a third aspect of the present invention, a processor is provided for running a program, wherein the program executes the cross-device fault diagnosis method based on heterogeneous signals as described above.

[0089] Example 2: The optional implementation process of the present invention is described below with reference to experimental data:

[0090] I. Implementation Environment

[0091] This embodiment is implemented in the following environment:

[0092] Hardware environment: A computer equipped with an Intel Core i9-13900 processor and 32GB of memory.

[0093] Software environment: Python 3.11

[0094] II. Implementation Steps

[0095] In this embodiment, a bearing test bench from Ottawa, Canada, was selected as the target equipment. The test bench consists of a single-phase motor mounted on a rigid plate, which is supported by a vibration damping bracket. The single-phase motor is raised using a shaft adapter, and its load end is equipped with an SKF E22206 spherical roller bearing to withstand the load applied by the cantilever beam, which is controlled by a lead screw. The motor is driven at a constant nominal speed of 1750 rpm, applying a constant load of 400 N. Two NSK 6203ZZ deep groove ball bearings at the drive end of the motor shaft are used as the test objects of the test bench; the bearing model at the drive end is changed after 5 tests; the first 5 tests used 6203ZZ deep groove ball bearings, and the subsequent 15 tests used FAFNIR 203KD deep groove ball bearings. The drive end bearing seals were removed, and the bearings were degreased to accelerate degradation during the test process. These bearings were reinstalled after each test. Data details are shown in Table 1 below.

[0096] Table 1

[0097]

[0098] Note: H-1-0, H represents health data, 1 represents bearing number 1, and 0 has no special meaning;

[0099] Note: B-11-1, B indicates rolling element fault data, 11 indicates bearing number 11, and 1 indicates minor fault;

[0100] Note: I-1-1, I represents inner ring fault data, 1 represents bearing number 1, and 1 represents minor fault;

[0101] Note: O-6-1, O indicates outer ring fault data, 6 indicates bearing number 6, and 1 indicates minor fault.

[0102] In this embodiment, based on the operating conditions of the target equipment, the Case Western Reserve University bearing test bench, which is compatible with both operating conditions and bearing type, is selected as the source equipment. The test bench consists of a 2-horsepower motor, a torque sensor / encoder, a dynamometer, and control electronics. The test object is the deep groove ball bearing at the motor fan end. A single-point fault is introduced through electrical discharge machining. Vibration data is collected using accelerometers mounted on a magnetic base on the motor housing. The accelerometers are positioned at the 12 o'clock position at both the motor drive end and the fan end. In this embodiment, 12kHz fan end bearing data is selected as the source device. The dataset is structured as shown in Table 2.

[0103] Table 2

[0104]

[0105] Note: Each row of data represents data under the same operating condition.

[0106] In this embodiment, a self-made bearing test bench with compatible operating conditions and bearing type is selected as the source equipment based on the target equipment's operating conditions. The self-made bearing test bench mainly consists of a three-phase asynchronous drive motor, a double-support bearing housing, a belt drive system, and a magnetic powder braking device. A 6205 type deep groove ball bearing was used as the test object. A normal control bearing was installed on the bearing housing near the motor, while a pre-set faulty bearing was installed on the other side. A single-point fault was introduced through electrical discharge machining. Vibration data was collected using an accelerometer horizontally and radially mounted on a magnetic bearing housing. In this embodiment, bearing data with a sampling rate of 25.6 kHz and a rotational speed of 2000 rpm was selected as the source device. The dataset contains three load conditions: 0 N.m, 3 Nm, and 6 N.m, and its specific composition is shown in Table 3.

[0107] Table 3

[0108]

[0109] According to the method of the present invention, the optional process of this embodiment is specifically described as follows:

[0110] 1) Based on the target equipment in Table 1, select the equipment in Tables 2 and 3 that are compatible with its operating conditions and have the same test object type as the source equipment, and deploy contact accelerometers on the rolling bearing measuring points of multiple source equipment. Collect vibration signals from the rolling bearings of multiple source equipment using the contact accelerometers, and simultaneously collect acoustic signals from the rolling bearings of the target equipment using non-contact microphones;

[0111] This embodiment uses 12kHz bearing data from the fan end of Case Western Reserve University as the source device. Vibration signal data, including 4 normal data points and 12 fault data points, with each data point lasting more than 10 seconds; a self-made test bench was used to acquire the data. Vibration signal data were sampled at a rate of 25.6 kHz, including 3 normal data points and 9 fault data points, with each data point lasting 60 seconds. The acoustic signal data of the bearings from the University of Ottawa, Canada, were selected as the target equipment data, with a sampling rate of 42 kHz, including 15 normal bearing data points and 3 fault data points, with each data point lasting 10 seconds.

[0112] Based on the acoustic operating conditions of the target device, with a rotational speed of 1750 rpm and a load of 400 N, vibration signals from multiple source devices that are compatible with the corresponding operating conditions are selected as multiple source domain vibration datasets for operating condition alignment, as shown in Tables 2 and 3. This step aims to ensure operating condition alignment.

[0113] 2) Based on the acoustic signal and the vibration signals in the multiple source domain vibration datasets, construct an acoustic feature dataset and a multi-source vibration feature dataset;

[0114] S21. For the acoustic signal and the vibration signals in the multiple source domain vibration datasets, a sliding window method with a window length of 1 second and an overlap rate of 10% is used for data augmentation to generate an augmented time-domain acoustic signal sample set and multiple source time-domain vibration signal sample sets for subsequent feature extraction. Specifically, the composition of the data augmentation sample set is shown in Table 4.

[0115] Table 4

[0116]

[0117] S23. Given that the difference in the original sampling frequency between the multi-source vibration signal and the acoustic signal results in different 1 / 3 octave band feature dimensions, in order to ensure effective transfer in knowledge distillation, the 1 / 3 octave band feature dimensions of each of the multiple source devices and the 1 / 3 octave band feature dimension of the target device are truncated.

[0118] In this embodiment of the invention, the source device The vibration signal sampling rate is 12kHz, with an effective frequency range of 0-6kHz. According to the IEC 61260 standard, its upper center frequency is defined as 4kHz, encompassing a 24-dimensional 1 / 3 octave band characteristic extending from 25Hz to 4kHz. Source device. The vibration signal sampling rate is higher at 25.6 kHz, with an effective frequency range of 0-12.8 kHz and an upper limit center frequency of 10 kHz, encompassing 28 dimensions of 1 / 3 octave band characteristics extending from 25 Hz to 10 kHz. The target domain acoustic signal sampling rate is 42 kHz, with an effective frequency range of 0-21 kHz. Based on the same standard, its upper limit frequency is determined to be 16 kHz, encompassing 30 dimensions of 1 / 3 octave band characteristics extending from 25 Hz to 16 kHz. To ensure the effectiveness of vibration knowledge transfer, this study uniformly uses the source device... Based on the 24 octave band characteristic dimensions, the source device The 1 / 3 octave band feature dimension of the source device and the 1 / 3 octave band feature dimension of the target device were truncated accordingly, retaining only the first 24 dimensions. The source device was obtained through the above process. Truncation of 1 / 3 octave band characteristics, source device The characteristics of the truncated 1 / 3 octave band and the characteristics of the target device after the truncated 1 / 3 octave band.

[0119] S24, Source Equipment Truncation of 1 / 3 octave band characteristics, source device The truncated 1 / 3 octave band characteristics of both the source device and the target device are preprocessed using a maximum-minimum normalization method to obtain their respective preprocessed 1 / 3 octave band characteristics; the source device... Source equipment The preprocessed 1 / 3 octave band features are integrated to construct a multi-source vibration feature dataset, and the preprocessed 1 / 3 octave band features of the target device are used to construct an acoustic feature dataset. The normalization formulas for maximum and minimum values ​​are as follows:

[0120] ;

[0121] Where x represents a 1 / 3 octave band characteristic; This represents the normalized 1 / 3 octave band characteristic.

[0122] 3) Randomly divide the vibration feature samples in the multi-source vibration feature dataset into the first training set and the first test set in an 8:2 ratio. Specifically, for the first division, the number of vibration feature samples is determined according to the floor rule. For the later division, the number of samples is the difference between the total number of samples and the number of samples in the first division.

[0123] 4) Based on the acoustic feature dataset and the multi-source vibration feature dataset, construct a small-sample acoustic-vibration paired training set, a second validation set, and a second test set;

[0124] Furthermore, S4 specifically includes:

[0125] S41. Construction of a small-sample acoustic training set: From the acoustic feature dataset constructed in step S24, 90% of the normal category samples and 1% of the fault category samples are extracted to form a small-sample acoustic training set. Specifically, based on the number of samples of the target device in Table 4, 90% of the normal category acoustic feature samples and 1% of the fault category acoustic feature samples are extracted from the acoustic feature dataset constructed in step S24 (the number of samples is determined by rounding down according to the proportion). It can be calculated that the small-sample acoustic training set contains 1228 normal category acoustic feature samples and 2 fault category acoustic feature samples. The severely unbalanced proportion is consistent with the typical situation of scarce fault samples in actual industrial monitoring scenarios.

[0126] S42. Construction of small sample acoustic-vibration pairing training set: Based on the small sample acoustic training set, randomly select M normal category labels and N fault category labels from the multi-source vibration feature dataset, and pair them one-to-one with the acoustic samples with the same label in the small sample acoustic training set to construct a small sample acoustic-vibration pairing training set with one-to-one correspondence between acoustic and vibration sample category labels.

[0127] For illustrative purposes, and not for limitation, let's take M=10 as an example. Assume the normal category samples in the small-sample acoustic training set are numbered 1-10, and the vibration feature samples with 10 normal category labels selected from the multi-source vibration feature dataset are numbered 11-20. Each number in the small-sample acoustic training set can only be paired with one number selected from the multi-source vibration feature dataset, and each number can only be paired once, ultimately forming 10 non-overlapping pairings. The pairing of fault category samples follows the same principle.

[0128] S43. After removing the small sample acoustic training set from the acoustic feature dataset, the remaining samples are randomly divided into a second validation set and a second test set according to the sample category labels in a ratio of 1:9.

[0129] Specifically: After removing the small sample acoustic training set from the acoustic feature dataset, the remaining normal samples are 137, and the faulty samples are 271. Following the rounding down principle, 10% of each of the normal and faulty samples are extracted as the second validation set, and the remainder is used as the second test set. That is, the second validation set contains 13 normal samples and 27 faulty samples; the second test set contains 124 normal samples and 244 faulty samples. The dataset structure of this embodiment is shown in Table 5.

[0130] Table 5

[0131]

[0132] 5) Construct a knowledge distillation framework, which includes a vibrating teacher model and a lightweight student model. The vibrating teacher model adopts 1D-ResNet-18, and its architecture is shown in Table 6. The lightweight student model adopts 1D-ResNet-6, and its architecture is shown in Table 7. The vibration teacher model is pre-trained and tested based on the first training set and the first test set to learn fault characterization knowledge of vibration signals and generate a pre-trained vibration teacher model. A lightweight student model is independently trained based on acoustic samples from a small-sample acoustic-vibration pairing training set to obtain a baseline student model without knowledge transfer. During knowledge distillation training on the small-sample acoustic-vibration pairing training set, in each epoch, the pre-trained vibration teacher model receives vibration samples from the small-sample acoustic-vibration pairing training set as input, while the lightweight student model receives acoustic samples with the same category label from the same small-sample acoustic-vibration pairing training set as input. The distillation loss is calculated by combining the temperature-softened soft labels output by the pre-trained vibration teacher model with the temperature-softened soft labels predicted by the lightweight student model. The hard loss is calculated by comparing the labels predicted by the lightweight student model with the true labels. Finally, a total loss function including distillation loss and hard loss is obtained. Validation is then performed on the second validation set, and hyperparameters (including temperature T, loss ratio, etc.) in the knowledge distillation process are optimized. Finally, we obtain an enhanced student model that has undergone knowledge distillation and transfer.

[0133] Table 6

[0134]

[0135] Table 7

[0136]

[0137] Generalization performance evaluation: The comprehensive performance evaluation is carried out by comparing the accuracy, F1 score and optimization process of the augmented student model and the independent trained benchmark student model on the second validation set, and combining the confusion matrix results on the second test set.

[0138] This embodiment combines 1 / 3 octave band features with knowledge distillation technology, utilizing bearing vibration signals from multiple source devices to guide the acoustic signal fault diagnosis knowledge of the target device, ultimately achieving enhanced anomaly detection based on acoustic signals under small sample and cross-device conditions. A schematic diagram of the accuracy and F1 score results based on the second validation set is shown below. Figure 2 As shown in the diagram, the confusion matrix results based on the second test set are illustrated below. Figure 3 As shown.

[0139] pass Figure 2It can be seen that, within 30 training cycles, when the target device was trained independently using only acoustic samples, the average accuracy of the model on the second validation set was 77.50%. However, after guiding acoustic diagnosis through knowledge transfer from vibration samples, the accuracy increased to 100.00%, representing an absolute performance gain of 22.5 percentage points. Similar results were also observed in the F1 score. These results fully demonstrate that the proposed transfer mechanism significantly enhances the anomaly detection and diagnostic performance of the student model for the target device.

[0140] Figure 3 The results further demonstrate that the small-sample acoustic-vibration paired training set constructed based on 1 / 3 octave band features resulted in only 2 sample recognition errors for the enhanced student model on the second test set; while the student model trained independently using only acoustic samples from the target device exhibited 74 sample recognition errors. This significant difference fully verifies the robustness of the method under small-sample conditions and effectively alleviates the challenge of scarce labeled data for target devices. These achievements fundamentally break through the accuracy bottleneck of non-contact acoustic monitoring, providing a reliable technical foundation for replacing contact sensors in industrial scenarios. Furthermore, the vibration teacher model has 15798 parameters, while the lightweight student model has only 358 parameters. The student model inferences 654 samples in just 589ms, with an average time of approximately 900ms per sample. It has the capability to be deployed on edge computing devices, which can meet the real-time needs of parallel monitoring of multiple devices in complex industrial scenarios.

[0141] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A cross-device fault diagnosis method based on heterogeneous signals, characterized in that, include: S1. Use a contact method to collect vibration signals of the test object on multiple source devices and construct multiple source domain vibration datasets; Acquire acoustic signals of the test object on the target device using a non-contact method; S2. Based on the acoustic signal and the vibration signals in the multiple source domain vibration datasets, construct an acoustic feature dataset and a multi-source vibration feature dataset; S3. The multi-source vibration feature dataset is randomly divided into a first training set and a first test set according to a first preset ratio; S4. Based on the acoustic feature dataset and the multi-source vibration feature dataset, construct a small-sample acoustic-vibration paired training set, a second validation set, and a second test set; S5. Construct a knowledge distillation framework, which includes a vibration teacher model and a lightweight student model; pre-train and test the vibration teacher model based on a first training set and a first test set to generate a pre-trained vibration teacher model; based on a small-sample acoustic-vibration paired training set and a second validation set, use the vibration knowledge extracted from the pre-trained vibration teacher model to guide the lightweight student model in training on acoustic signals, thereby realizing vibration knowledge transfer; S2 specifically includes: S21. The acoustic signal and the vibration signals in the multiple source domain vibration datasets are respectively augmented using the sliding window method to generate an augmented time-domain acoustic signal sample set and multiple source time-domain vibration signal sample sets. S22. Within a preset frequency band, extract 1 / 3 octave band features on the corresponding center frequency point bandwidth from the samples in the enhanced time-domain acoustic signal sample set and the samples in the multiple source time-domain vibration signal sample sets. S23. Perform feature dimension truncation processing on the 1 / 3 octave band feature dimension of each of the multiple source devices and the 1 / 3 octave band feature dimension of the target device to obtain the truncated 1 / 3 octave band features of the multiple source devices and the truncated 1 / 3 octave band features of the target device. S24. Preprocess the truncated 1 / 3 octave band features of the multiple source devices and the truncated 1 / 3 octave band features of the target device to obtain preprocessed 1 / 3 octave band features; integrate the preprocessed 1 / 3 octave band features of the multiple source devices to construct a multi-source vibration feature dataset, and construct an acoustic feature dataset from the preprocessed 1 / 3 octave band features of the target device. S5 specifically includes: S51. Constructing a knowledge distillation framework: The vibration teacher model uses ResNet-18 as its basic architecture, and performs dimensionality reduction transformation on the convolutional layers in ResNet-18 from two dimensions to one dimension: replacing the two-dimensional convolutional kernels with one-dimensional convolutional kernels and reducing the channel dimension of the convolutional layers; through this transformation, a lightweight one-dimensional improved architecture 1D-ResNet-18 is constructed as the vibration teacher model; the lightweight student model is based on the lightweight one-dimensional improved architecture 1D-ResNet-18, and by reducing the number of residual blocks to 1 and introducing a max pooling layer after the residual blocks, a lightweight student model 1D-ResNet-6 is constructed; S52. Vibration teacher model pre-training: Based on the first training set and the first test set, the vibration teacher model is pre-trained and tested to generate a pre-trained vibration teacher model. S53. During knowledge distillation training on a small sample acoustic-vibration pairing training set, in each epoch of training, the pre-trained vibration teacher model receives vibration samples from the small sample acoustic-vibration pairing training set as input, while the lightweight student model receives acoustic samples with the same category label from the small sample acoustic-vibration pairing training set as input. The distillation loss is formed by the temperature-softened soft label output by the pre-trained vibration teacher model and the temperature-softened soft label predicted by the lightweight student model. The hard loss is calculated by comparing the label predicted by the lightweight student model with the true label. Finally, the total loss function including the distillation loss and the hard loss is obtained. Then, the model is validated on the second validation set, and the hyperparameters in the knowledge distillation process are optimized to obtain the enhanced student model after knowledge distillation transfer.

2. The cross-device fault diagnosis method based on heterogeneous signals according to claim 1, characterized in that, S1 includes: S11. Select a device that is compatible with the operating conditions of the target device and has the same type of test object as the source device; and deploy contact acceleration sensors on the test objects of multiple source devices. S12. Vibration signals of multiple source devices are collected from the test object using a contact accelerometer; acoustic signals of the target device are collected from the test object using a non-contact microphone. S13. Based on the operating condition parameters of the target equipment, select a subset of data that matches the operating conditions from the vibration signals of multiple source equipment, and construct multiple source domain vibration datasets that are aligned with the operating conditions.

3. The cross-device fault diagnosis method based on heterogeneous signals according to claim 2, characterized in that, The adaptation to the target equipment's operating conditions specifically refers to: Using rotational speed as the target equipment's operating condition parameter, select the equipment whose operating condition parameter value differs from the target equipment's value within a preset threshold as the source equipment.

4. The cross-device fault diagnosis method based on heterogeneous signals according to claim 1, characterized in that, Specifically, S23 is as follows: The minimum value of the 1 / 3 octave band characteristic dimension among the multiple source devices is taken as the first dimension value; the source device with the 1 / 3 octave band characteristic dimension value as the first dimension value is taken as the first source device, and the rest are second source devices; the 1 / 3 octave band characteristic dimension value of the target device is taken as the second dimension value. The first dimension value and the second dimension value are compared: if the first dimension value is less than the second dimension value, the 1 / 3 octave band feature dimension of the second source device and the target device is truncated so that the truncated 1 / 3 octave band feature dimension value is the first dimension value; otherwise, the 1 / 3 octave band feature dimension of the second source device is truncated so that the truncated 1 / 3 octave band feature dimension value is the first dimension value.

5. The cross-device fault diagnosis method based on heterogeneous signals according to claim 1, characterized in that, The S4 includes: S41. Construction of a small-sample acoustic training set: From the acoustic feature dataset, M normal category acoustic feature samples with a first preset proportion and N fault category acoustic feature samples with a second preset proportion are extracted according to the floor rule to form a small-sample acoustic training set; wherein, the first preset proportion is greater than the second preset proportion. S42. Construction of small sample acoustic-vibration pairing training set: Based on the small sample acoustic training set, randomly select M normal category labels and N fault category labels from the multi-source vibration feature dataset, and pair them one by one with the acoustic samples with the same label in the small sample acoustic training set to construct the small sample acoustic-vibration pairing training set. S43. The samples in the acoustic feature dataset, excluding the small sample acoustic training set, are randomly divided into a second validation set and a second test set according to the sample category labels and a second preset ratio.

Images