A transfer learning method for classification based on one-dimensional periodic signals

Abstract

The present invention provides a transfer learning method for classification based on one-dimensional periodic signals, comprising: S1, using an encoding neural network to extract a feature vector of each sample in a source domain data set and a target domain data set, using the feature vectors and labels of samples with labels in the source domain and the target domain to train a classifier to identify the state category of a device, and updating the parameters of the classifier according to the classification loss during training; S2, inputting the feature vector of each unlabeled sample in the target domain into the classifier trained in step S1 to output the state category corresponding to the sample, and using it as a pseudo label of the corresponding sample; S3, determining a first joint distribution loss between the source domain and the target domain, which includes marginal distribution differences and conditional distribution differences, based on the feature vectors and labels of the samples in the source domain data set and the feature vectors, labels and pseudo labels of the samples in the target domain data set, and updating the parameters of the encoding neural network according to the first joint distribution loss.

Description

Technical Field

[0001] The present invention relates to the field of neural networks, specifically to the field of transfer learning, and more specifically to a transfer learning method for classification based on one-dimensional periodic signals. Background Art

[0002] In industry, various devices exhibit different performance in different operating states (e.g., normal and various fault conditions). Sensors installed on the devices or deployed on them can collect one-dimensional periodic signals from these devices. Different one-dimensional periodic signals are collected under different device states. A one-dimensional periodic signal is vector data collected by sensors that periodically changes over time. By labeling the collected one-dimensional periodic signals with labels indicating the device's status, a dataset is created and used to train a classification model. The trained classification model can then be used to monitor device status, allowing users to easily understand the device's status.

[0003] Industrial equipment comes in many different types and models. Due to differences in design, parameters, materials, and structure, the distribution of one-dimensional periodic signals collected from these devices often differs. After training a classification model for monitoring a particular device based on a labeled dataset, the trained model cannot be directly applied to monitoring other devices. However, creating a dataset with numerous labeled samples for each device would be labor-intensive and time-consuming. Therefore, transfer learning techniques can be used to reduce these costs.

[0004] Existing transfer learning technologies for one-dimensional periodic signals mainly include the following two:

[0005] The first technology: a unified feature extraction method is used for the one-dimensional periodic signal sets in the source domain and the target domain. The means are not limited to physical quantities and deep learning methods. Then, model training is used to shorten the distribution difference between the two feature sets to achieve the purpose of migration. For example, the related technology can be found in the Chinese patent application with publication number CN113869563A. In this application, the features output by the last fully connected layer of the classification neural network are used for migration. Although this migration method achieves the purpose of migration, it fails to enrich the distribution of the original signal in the target domain. The physical meaning is also eliminated in the process of feature extraction, resulting in a decrease in interpretability. In addition, it only uses the root mean square error between features when evaluating the migration effect. Although small numerical differences in dimensionless normalized features can be approximated to small distribution differences, this cannot be approximated in other scenarios. It has poor generalization and lacks persuasiveness in other signal scenarios.

[0006] The second technique involves pre-training a single model using a large amount of source domain data to achieve good results on the source domain data. The pre-trained model is then fine-tuned ("finetuning") using a small amount of target domain data to achieve transfer. In other words, the transfer method used by the second technique involves training a pre-trained prediction model in the source domain and then fine-tuning the parameters of the pre-trained prediction model in the target domain based on small-sample training. If the conditional and marginal distributions of the source and target domains differ, this technique will struggle to achieve good transfer results and will also suffer from poor generalization and lack of persuasiveness in other signal scenarios.

[0007] In general, when detecting the status of a device based on a one-dimensional periodic signal, the following problems exist:

[0008] 1. Existing technology that uses pre-trained prediction models to achieve transfer learning requires that the neural network be trained on source domain data to establish network parameters. Then, by retraining some layers of the neural network with small sample data from the target domain, these parameters are modified to achieve the transfer objective. However, this approach has a problem: the small sample from the target domain cannot replicate the training parameters of the large sample from the source domain, negating the effectiveness of pre-training. This makes the transfer effect highly dependent on the distributional similarity between the source and target domains. If the conditional and marginal distributions of the two domains differ significantly, this transfer method is virtually ineffective.

[0009] 2. If you want to directly apply mature image migration technology to the one-dimensional signal field, you must use some methods to convert the one-dimensional vector data into two-dimensional matrix data. However, this approach has many problems, such as: 1) The network structure is not suitable for the feature extraction method of one-dimensional signals, and the convolution kernel structure causes significant loss of original information during feature extraction; 2) During data processing, many one-dimensional signals are converted into two-dimensional matrices after applying mathematical formulas. The converted matrices have no physical meaning and lose their application interpretability; 3) The computational cost increases exponentially.

[0010] Therefore, it is necessary to improve the existing migration technology based on one-dimensional periodic signals to adapt to various application scenarios. Summary of the Invention

[0011] Therefore, the purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a transfer learning method for classification based on one-dimensional periodic signals.

[0012] The purpose of the present invention is achieved through the following technical solutions:

[0013] According to a first aspect of the present invention, a transfer learning method for classification based on one-dimensional periodic signals is provided, the method comprising the following steps: S1, extracting a feature vector of each sample in a source domain dataset and a target domain dataset using an encoding neural network, training a classifier using the feature vectors and labels of labeled samples in the source domain and the target domain to identify the state category of a device, and updating the parameters of the classifier according to the classification loss, wherein the samples in the source domain and the target domain are collected from different devices, all samples in the source domain are labeled, and some samples in the target domain are labeled, each label indicates the state category corresponding to the device when the corresponding sample is collected, and the sample is a one-dimensional periodic signal collected for the device; S2, inputting the feature vector of each unlabeled sample in the target domain into the classifier trained in step S1 to output the state category corresponding to the sample, and using it as a pseudo label for the corresponding sample; S3, determining a first joint distribution loss between the source domain and the target domain, including marginal distribution difference and conditional distribution difference, based on the feature vectors and labels of the samples in the source domain dataset and the feature vectors, labels and pseudo labels of the samples in the target domain dataset, and updating the parameters of the encoding neural network according to the first joint distribution loss.

[0014] In some embodiments of the present invention, the method further includes: S4, using the encoding neural network whose parameters are updated in step S3 to extract the feature vector of each sample in the source domain dataset, and using the decoding neural network to construct a pseudo sample of the target domain based on the obtained feature vector of each sample to obtain a pseudo target domain dataset; judging whether the transfer learning meets the predetermined requirements based on the classification loss, the first joint distribution loss, and the comprehensive loss corresponding to multiple distribution differences between the pseudo target domain dataset and the target domain fully labeled dataset with the pseudo labels obtained in step S3, and continuing to perform transfer learning if the predetermined requirements are not met.

[0015] In some embodiments of the present invention, after updating the parameters of the encoding neural network using all samples in the source domain dataset and the target domain dataset each time in step S3, step S4 is performed while locking the parameters of the encoding neural network, and when transfer learning needs to be continued, the parameters of the encoding neural network are unlocked to continue transfer learning.

[0016] In some embodiments of the present invention, the step S4 includes: S41, using a decoding neural network to construct a pseudo sample of the target domain based on the feature vector of each sample in the source domain extracted by the encoding neural network whose parameters are updated in step S3, to obtain a pseudo target domain dataset; S42, decomposing the target domain fully annotated dataset and the pseudo target domain dataset into feature sets with different information, and calculating the comprehensive loss corresponding to the distribution difference between the target domain fully annotated dataset and the pseudo target domain dataset, and between the feature sets corresponding to the decomposition of the target domain fully annotated dataset and the pseudo target domain dataset; S43, judging whether the classification loss, the first joint distribution loss and the comprehensive loss have all converged, if not, continuing to perform transfer learning.

[0017] In some embodiments of the present invention, the target domain fully annotated dataset and the pseudo-target domain dataset are respectively decomposed into the following feature sets or their combinations: a target domain periodic feature set with periodic information and a pseudo-target domain periodic feature set; a target domain trend feature set with trend information and a pseudo-target domain trend feature set; a target domain period multiple feature set and a pseudo-target domain period multiple feature set with period multiple information.

[0018] In some embodiments of the present invention, the comprehensive loss includes: a second joint distribution loss including marginal distribution differences and conditional distribution differences between the target domain fully annotated dataset and the pseudo target domain dataset, a third joint distribution loss including marginal distribution differences and conditional distribution differences between the target domain periodic feature set and the pseudo target domain periodic feature set, cosine similarity between the target domain trend feature set and the pseudo target domain trend feature set, a fourth joint distribution loss including marginal distribution differences and conditional distribution differences between the target domain period multiple feature set and the pseudo target domain period multiple feature set, or a combination thereof, wherein the loss included in the comprehensive loss is related to the feature sets decomposed from the target domain fully annotated dataset and the pseudo target domain dataset respectively.

[0019] In some embodiments of the present invention, the encoding neural network is a down-sampled one-dimensional convolutional neural network, which is used to perform multiple batch normalization, convolution processing and down-sampling processing on the sample data, and then output the feature vector of the sample through global average pooling processing.

[0020] In some embodiments of the present invention, the encoding neural network includes multiple convolutional layers for convolution processing, wherein the width of the convolution kernel of the first convolution layer is set according to the number of correlations of the signal source of the input one-dimensional periodic signal, and the length of the convolution kernel of the first convolution layer is the least common multiple of the period lengths of all one-dimensional periodic signals of the source domain dataset and the target domain dataset; and the size of the convolution kernel of the convolution layer after the first convolution layer is selected from a plurality of preset convolution kernel sizes, and the size of the convolution kernel of the convolution layer after the first convolution layer is equal to or smaller than the size of the convolution kernel of the convolution layer of the previous layer.

[0021] In some embodiments of the present invention, the decoding neural network is an up-sampled one-dimensional convolutional neural network, which is used to perform multiple deconvolution processes and up-sampling processes on the feature vectors of samples in the source domain and output pseudo samples in the target domain.

[0022] According to a second aspect of the present invention, a method for classifying device status is provided, comprising: obtaining a one-dimensional periodic signal collected from the device and using it as a sample to be predicted; inputting the sample to be predicted into an encoding neural network and classifier obtained by the method described in the first aspect, and outputting the status category of the device.

[0023] According to a third aspect of the present invention, there is provided a transfer learning system for classification based on one-dimensional periodic signals, comprising: an encoding neural network and a classifier, wherein the transfer learning system is configured to train the encoding neural network and the classifier according to the method of the first aspect.

[0024] In some embodiments of the present invention, the encoding neural network is a down-sampled one-dimensional convolutional neural network, which is used to perform multiple batch normalization, convolution processing and down-sampling processing on the sample data, and then output the feature vector of the sample through global average pooling processing; the classifier is a fully connected network classifier, which identifies the status category of the device based on the feature vector of the sample.

[0025] In some embodiments of the present invention, the transfer learning system also includes a decoding neural network, which is an upsampled one-dimensional convolutional neural network, which is used to perform multiple deconvolution and upsampling processes on the feature vectors of samples in the source domain and output pseudo samples of the target domain.

[0026] According to a fourth aspect of the present invention, an electronic device is provided, comprising: one or more processors; and a memory, wherein the memory is used to store executable instructions; the one or more processors are configured to implement the steps of the method described in the first aspect and / or the second aspect by executing the executable instructions.

[0027] Compared with the prior art, the advantages of the present invention are:

[0028] The present invention can train the encoding neural network during transfer learning to reduce the marginal distribution and conditional distribution between the data in the source domain and the target domain, better transfer the knowledge in the source domain to the target domain, and improve the accuracy of the model in identifying the state category of the device;

[0029] The present invention judges whether the transfer learning reaches the predetermined requirements based on the comprehensive losses corresponding to multiple distribution differences between the pseudo target domain dataset constructed by the decoding neural network and the fully annotated target domain dataset. It can enable the encoding neural network to better learn the knowledge of reducing the distribution differences between the source domain and the target domain, and better improve the accuracy of classification based on the feature vector extracted by the encoding neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The embodiments of the present invention are further described below with reference to the accompanying drawings, in which:

[0031] Figure 1 is a schematic diagram of a schematic encoding neural network according to an embodiment of the present invention;

[0032] Figure 2 is a schematic diagram of a decoding neural network according to an embodiment of the present invention;

[0033] Figure 3 Schematic diagram of the process of transfer learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0035] As mentioned in the background technology section, if the existing technology wants to use a pre-trained prediction model to achieve transfer learning, its neural network must first be trained with source domain data to form network parameters, and then some parameters are changed by retraining some layers of the neural network with small sample data from the target domain to finally achieve the purpose of migration. However, the effect of migration is highly dependent on the distribution similarity between the source domain and the target domain. If the conditional distribution and marginal distribution of the two are very different, such a migration method can be said to be almost ineffective. If you directly borrow mature image migration technology, you will inevitably need to use some methods to convert one-dimensional vector data into two-dimensional matrix data. However, this method has problems such as the convolution kernel structure causing a large loss of original information during feature extraction, poor interpretability, and exponential growth in computational costs. To address the above-mentioned problems, the present invention provides a transfer learning method for classification based on one-dimensional periodic signals. In summary, the present invention mixes samples from the source domain and samples from the target domain (it should be understood that, for the sake of simplicity, the present invention sometimes abbreviates samples of the source domain dataset as source domain samples, and samples of the target domain dataset as target domain samples) for training an encoding neural network and a classifier, and uses the feature vectors and labels of labeled samples in the source and target domains to train the classifier to identify the state category of the device. During training, the parameters of the classifier are updated according to the classification loss; the parameters of the encoding neural network are updated by determining the first joint distribution loss between the source and target domains based on the feature vectors and labels of samples in the source domain dataset and the feature vectors, labels, and pseudo-labels of samples in the target domain dataset, so as to narrow the marginal distribution and conditional distribution of samples in the source and target domains; thus, compared with the prior art, the present invention can train the encoding neural network to narrow the marginal distribution and conditional distribution between the data in the source and target domains during transfer learning, better transfer the knowledge of the source domain to the target domain, and improve the accuracy of the model in identifying the state category of the device.

[0036] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0037] Before describing the embodiments of the present invention in detail, some of the terms used therein are explained as follows:

[0038] Transfer learning, also known as inductive transfer or domain adaptation, aims to apply knowledge or patterns from a certain domain (source domain) to a different but related domain (target domain) or problem.

[0039] The source domain refers to the domain where the existing knowledge in transfer learning is located.

[0040] The target domain refers to another domain to which the knowledge of the source domain needs to be transferred in transfer learning.

[0041] Marginal distribution refers to the probability distribution obeyed by part of the data in the probability distribution of multidimensional data.

[0042] Conditional distribution refers to the posterior probability distribution between data and labels. Conditional distribution is related to the learning task.

[0043] The following is an explanation from four aspects: model structure design, data set acquisition, transfer learning, and application scenarios.

[0044] 1. Design of model structure

[0045] In order to predict and classify the status of a device using a one-dimensional periodic signal, a corresponding prediction model needs to be established. According to one embodiment of the present invention, the prediction model includes an encoding neural network and a classifier. The encoding neural network can adopt a one-dimensional convolutional neural network (1D-CNN), such as a known one-dimensional convolutional neural network. The classifier can adopt a fully connected network classifier. According to one embodiment of the present invention, the encoding neural network is a downsampled one-dimensional convolutional neural network that performs multiple batch normalization (BN), convolution, and downsampling processes on the sample data, and then outputs the sample's feature vector through global average pooling.

[0046] Furthermore, to make the encoding neural network structure more suitable for processing one-dimensional periodic signals and improve the model's prediction accuracy, the encoding neural network of the present invention includes multiple convolutional layers for convolution processing. The width of the convolution kernel of the first convolutional layer is set based on the number of correlations (N) among the signal sources of the input one-dimensional periodic signal. The length of the convolution kernel of the first convolutional layer is the least common multiple of the period lengths of all one-dimensional periodic signals in the source and target domain datasets (i.e., the least common multiple of the period lengths of the one-dimensional periodic signals from all signal sources, represented by L). Therefore, if the input one-dimensional periodic signal includes correlated signals, the convolution kernel size of the first convolutional layer can be set to L×N. Regarding the convolution kernel width, assuming that the device has M sensors (the number of signal sources is represented by M, e.g., M=4 for four signal sources), and the signals of every two signal sources are correlated (i.e., the number of correlations is 2), the width of the convolution kernel of the first convolutional layer is 2, i.e., N=2. For example, suppose a piece of machinery is equipped with four vibration sensors, where M = 4. Sensors 1 and 2 are placed on one bearing seat, and sensors 3 and 4 are placed on another bearing seat. Therefore, the number of sensors with cross-correlation on each bearing seat is 2, or N = 2. However, it should be understood that the number of correlations here can be set according to the needs of the implementer. During implementation, the implementer can continuously experiment with adjusting the value of N. The implementer can also treat each signal as independent, setting the number of correlations to 1, and then setting the width of the convolution kernel of the first convolution layer to 1. Alternatively, the implementer can set N to 2 or other values ​​to observe the accuracy of the final prediction model and select the optimal setting. Regarding the length of the convolution kernel, assuming that the period lengths of the signals collected by the four sensors are 2, 4, 5, and 2, respectively, the least common multiple of the period lengths of the one-dimensional periodic signals of all signal sources is 20. The length of the convolution kernel in the first convolution layer of the encoding neural network is set to 20. Furthermore, the convolution kernel sizes of convolution layers after the first convolution layer are selected from a plurality of preset convolution kernel sizes, and the sizes of the convolution kernels of convolution layers after the first convolution layer are equal to or smaller than the sizes of the convolution kernels of the previous convolution layer. For example, the width of the convolution kernels of convolution layers after the first convolution layer is 1, and the length of the convolution kernels of convolution layers after the first convolution layer is selected from a plurality of preset convolution kernel sizes, and the length of the convolution kernels of convolution layers after the first convolution layer is equal to or smaller than the length of the convolution kernels of the previous convolution layer. For example, except for the first convolution layer, the convolution kernels of the other convolution layers are selected from Gaussian wavelet convolution kernels with preset convolution kernel sizes of 3×1, 5×1, 7×1, and 9×1, respectively. Each convolution layer is constructed by keeping the convolution kernel sizes constant or decreasing layer by layer, and the horizontal and vertical movement step size of the convolution kernels of each layer is [2, 1]. Pooling layers are provided between convolution layers to extract more abstract features.It should be understood that the above settings of the convolution kernel size and moving step are only illustrative, and those skilled in the art can set them as needed during implementation, and the present invention does not impose any limitations on this.

[0047] Furthermore, according to one embodiment of the present invention, the encoding neural network is a down-sampled one-dimensional convolutional neural network, and the encoding neural network for encoding the data of the input sample includes: multiple intermediate units and an output layer, wherein: each intermediate unit includes: a BN layer for performing batch normalization processing on the data input to this layer; a convolution layer for performing convolution processing on the data input to this layer; and a down-sampling layer for performing pooling processing on the data input to this layer. The output layer is used to perform global average pooling on the data input to this layer and output the feature vector of the corresponding sample. It should be understood that in this field, the number of intermediate units can be customized as needed, and the present invention does not impose any restrictions on this. According to one embodiment of the present invention, Figure 1 A schematic diagram of the structure of a schematic encoding neural network is shown, wherein: Figure 1 a is the size (specification) of the input data of the encoding neural network, the intermediate data input to or output from the corresponding layer, and the output data, Figure 1 The four dimensions in the brackets of a from left to right represent the number of samples, data width, data length, and number of data channels, respectively; among them, C is the number of input data (the number of samples); L is the common multiple of the period of the one-dimensional periodic signal (the least common multiple of the period lengths of the one-dimensional periodic signals of all signal sources in each sample); N is the number of correlated signals in the M one-dimensional periodic input signals (for example, a mechanical device is equipped with a total of 4 vibration sensors, then M=4, where sensors 1 and 2 are placed on one bearing seat, and sensors 3 and 4 are placed on another bearing seat, then the number of correlated sensors on each bearing seat is 2, N = 2); LEN is the sampling length of the one-dimensional periodic signal. For example, according to the above definition, (C, M, LEN, 1) at the input data indicates that the number of input samples is C, the width of the data is M (each sample contains M one-dimensional periodic signals), the length of the data is LEN (the length of each one-dimensional periodic signal is LEN), and the number of channels is 1; (C, M, LEN, 1) at the intermediate data indicates that the number of samples is C, the width of the data is M, the length of the data is LEN, and the number of channels is 1; (C, 64) at the output data indicates the feature vectors of the output C samples, and the number of channels of the feature vector of each sample is 64; Figure 1b is a schematic diagram of the encoding neural network structure, where the first layer is a batch normalization layer; the second layer is a convolutional layer (convolution kernel L×N; stride [L / 2, N]); the third layer is a maximum pooling layer ([2, 1]); the fourth layer is a batch normalization layer; the fifth layer is a convolutional layer (convolution kernel 7*1; stride [2, 1]); the sixth layer is a maximum pooling layer ([2, 1]); the seventh layer is a batch normalization layer; the eighth layer is a convolutional layer (convolution kernel 3*1; stride [2, 1]); the ninth layer is a maximum pooling layer ([2, 1]); and the tenth layer is a global average pooling layer. This encoding neural network has three intermediate units, corresponding to the intermediate units of layers 1-3, layers 4-6, and layers 7-9, respectively. The pooling layer corresponds to the downsampling layer, and the global average pooling layer corresponds to the output layer. Among them, the convolution layer (convolution kernel L×N) indicates that the length of the convolution kernel is L and the width is N; the step size [L / 2,N]) indicates that the lateral movement step size of the convolution kernel is L / 2, and the vertical movement step size is N; the maximum pooling layer ([2,1] indicates that the lateral movement step size of the pooling processing kernel during pooling is 2, and the vertical movement step size is 1. The rest is similar and will not be elaborated here. It should be understood that the number of intermediate units can be adaptively increased or decreased according to the specific classification task to be predicted; the size of the convolution kernel, the step size, the pooling method of the pooling layer, etc. can also be adjusted to other pooling methods, and the present invention does not impose any restrictions on this.

[0048] According to one embodiment of the present invention, if it is necessary to better supervise whether the transfer learning meets the predetermined requirements, a decoding neural network can be set to supervise the transfer learning of the encoding neural network. The decoding neural network can extract the feature vector of each sample in the source domain data set based on the encoding neural network to construct a pseudo sample of the target domain (equivalent to the size of the input data of the decoding neural network being the size of the output data of the encoding neural network, and the size of the output data of the decoding neural network being the size of the input data of the encoding neural network), and further use the pseudo sample of the target domain to supervise the transfer learning process of the encoding neural network. According to one embodiment of the present invention, the decoding neural network is an upsampled one-dimensional convolutional neural network, which is used to perform multiple deconvolution and upsampling processes on the feature vectors of the samples in the source domain, and then output the pseudo sample of the target domain. The decoding neural network includes: multiple intermediate units, wherein: each intermediate unit includes: a deconvolution layer, which is used to perform deconvolution processing on the data input to this layer, and an upsampling layer, which is used to perform upsampling processing on the data input to this layer; the data output by the last upsampling layer is the output data. According to one embodiment of the present invention, Figure 2 is a schematic diagram showing the structure of a schematic decoding neural network, wherein, Figure 2 a represents the input data of the decoding neural network, the intermediate data input to or output from the corresponding layer, and the output data specifications. The meaning of each letter is the same as that in the encoding neural network and is not repeated here; Figure 2b is a schematic diagram of the structure of the decoding neural network, where the first layer: deconvolution layer (deconvolution [M / N, LEN / (L / 2) / 32]); the second layer: upsampling layer ([2, 1]); the third layer: deconvolution layer (deconvolution [M / N, LEN / (L / 2) / 8]); the fourth layer: upsampling layer ([2, 1]); the fifth layer: deconvolution layer (deconvolution [M / N, LEN / (L / 2) / 2]); the sixth layer: upsampling layer ([L, N]). Among them, deconvolution [M / N, LEN / (L / 2) / 32] means that the input data is deconvolved through the inverse convolutional neural network operation to output a feature vector of size [C, M / N, LEN / (L / 2) / 32, 64]; the upsampling layer ([L, N]) means upsampling the feature vector output of the previous layer, wherein the first value in [L, N] represents upsampling the eigenvalue of the input in the length direction (the second to last dimension in the size), and the upsampling multiple is L (one eigenvalue in the length direction becomes L), and the second value represents upsampling the eigenvalue of the input in the width direction (the third to last dimension in the size), and the upsampling multiple is N (one eigenvalue in the length direction becomes N). The rest is similar and will not be described here. It should be understood that the structure of the decoding neural network is only schematic. The purpose of the decoding neural network is to restore the size of the data from the size of the feature vector to the size of the sample. The actual structure can be set as needed according to the size of the input feature vector and the size of the sample. The present invention does not impose any restrictions on this.

[0049] 2. Dataset Acquisition

[0050] To perform transfer learning, a source domain dataset and a target domain dataset must be prepared in advance. According to one embodiment of the present invention, the source domain dataset includes multiple samples. These samples are one-dimensional periodic signals collected from a device. All source domain samples are labeled, with each label indicating the device's state category at the time the corresponding sample was collected. The target domain dataset includes multiple samples, also one-dimensional periodic signals collected from a target device. Some target domain samples have corresponding labels, while others do not. The labels indicate the target device's state category at the time the corresponding sample was collected. Furthermore, the source and target domain samples are collected from different devices. For example, the source domain samples are collected from a first device (e.g., a mechanical fault simulator), while the target domain samples are collected from a second device (e.g., a pump). According to one embodiment of the present invention, the state categories indicated by the sample labels include normal states and multiple fault states. (For example, assuming the state categories of a mechanical device with rotating parts need to be monitored, the possible fault states include imbalance, misalignment, deflection, looseness, and bearing damage. It should be understood that this is for illustrative purposes only. Implementers can configure specific fault categories for other types of devices as needed, and the present invention does not impose any limitations on this.) The number of condition categories indicated by the labels of the source domain samples (for example, five categories: imbalance, misalignment, deflection, looseness, and bearing damage) is greater than or equal to the number of condition categories indicated by the labels of the target domain (for example, three categories: imbalance, misalignment, and looseness). For example, in the source domain dataset, all samples are labeled with one of the following: normal, imbalance, misalignment, deflection, looseness, and bearing damage; while in the target domain dataset, some samples are labeled with at least one of the following: normal, imbalance, misalignment, and looseness. Alternatively, although the ultimate goal for the target domain is to predict the four condition categories of normal, imbalance, misalignment, and looseness, when labeling the target domain samples, at least a portion of the data belonging to each condition category can be labeled, or only a portion of the samples belonging to some condition categories can be labeled, leaving the rest unlabeled. For example, only a portion of the samples belonging to normal can be labeled, or a portion of the samples belonging to both normal and imbalanced can be labeled. It should be understood that the status categories of the aforementioned devices are only illustrative. When implementing on different devices, those skilled in the art can set the status categories of the devices that may be involved as needed, and the present invention does not impose any restrictions on this. For example, in another embodiment, the status categories indicated by the labels of the samples in the source domain include normal, loose nuts, and broken bearings; while the status categories indicated by the labels of the samples in the target domain include normal and loose nuts.

[0051] 3. Transfer Learning

[0052] In order to transfer the knowledge of the source domain dataset to the target domain, transfer learning is required. According to one embodiment of the present invention, the transfer learning process includes:

[0053] Step S1: Use an encoding neural network to extract the feature vector of each sample in the source domain and the target domain, and use the feature vectors and labels of the samples with labels in the source domain and the target domain to train a classifier to identify the status category of the device. During training, the parameters of the classifier are updated according to the classification loss.

[0054] According to one embodiment of the present invention, samples in the source domain and the target domain are collected from different devices. All samples in the source domain are labeled, and some samples in the target domain are labeled. Each label indicates the state category of the device when the corresponding sample is collected. The sample is a one-dimensional periodic signal collected from the device. Figure 3 The samples in the source domain dataset A (all data are labeled, i.e., all samples have corresponding labels) and the target domain dataset B (partial data are labeled, i.e., some samples have corresponding labels) are input into the encoding neural network, which outputs the feature vector of each sample in the source domain and the feature vector of each sample in the target domain. The feature vectors of all source domain samples constitute the source domain feature vector set a; the feature vectors of all target domain samples constitute the target domain feature vector set b (which includes the target domain labeled feature vector set b-labeled and the unlabeled feature vector set b-unlabeled). The source domain feature vector set a and the label of each sample, as well as the feature vectors of the labeled samples in the target domain feature vector set b and their labels (the target domain labeled feature vector set b-labeled) constitute the classifier training set. The classifier is trained using the classifier training set to identify the status category of the device. During training, the classifier parameters are updated based on the classification loss.

[0055] Step S2: Input the feature vector of each sample without a label in the target domain into the classifier trained in step S1 to output the state category corresponding to the sample, and use it as the pseudo label of the corresponding sample;

[0056] In some embodiments of the present invention, still refer to Figure 3 The feature vector of each sample in the target domain's unlabeled feature vector set b-unlabeled is input into the classifier trained in step S1 to output the corresponding state category of the sample, which is used as the pseudo-label for the corresponding sample. The calculation of the conditional distribution difference depends on the feature vector and the label. Since only some samples in the target domain have labels, obtaining the pseudo-label facilitates the subsequent calculation of the conditional distribution difference between the current source and target domains.

[0057] Step S3: Determine the first joint distribution loss between the source domain and the target domain, which includes marginal distribution differences and conditional distribution differences, based on the feature vectors and labels of the samples in the source domain dataset and the feature vectors, labels, and pseudo-labels of the samples in the target domain dataset, and update the parameters of the encoding neural network based on the first joint distribution loss.

[0058] According to one embodiment of the present invention, the marginal distribution difference between the source domain and the target domain is obtained by calculating the marginal distribution difference between the feature vectors of the samples in the source domain dataset (corresponding to the source domain feature vector set a) and the feature vectors of the samples in the target domain dataset (corresponding to the target domain feature vector set b) (for example, by calculating the MK-MMD distance. The marginal distribution difference is calculated in a similar manner elsewhere in the present invention and will not be described in detail later). Figure 3 In this example, the full label feature vector set bla refers to the full label feature vector set of the source domain and the full label feature vector set of the target domain. The conditional distribution difference between the source and target domains is calculated by computing the conditional distribution difference between the feature vectors and labels of the samples in the source domain dataset (the source domain full label feature vector set) and the feature vectors, labels, and pseudo-labels of the samples in the target domain dataset (the target domain full label feature vector set). The first joint distribution loss is obtained by summing or weighting the marginal distribution difference and the conditional distribution difference between the source and target domains. The gradient of the first joint distribution loss is calculated, and the parameters of the encoding neural network are updated based on the gradient backpropagation.

[0059] Through the transfer learning of steps S1-S3, the knowledge of the source domain dataset can be transferred to the target domain. In addition, in order to obtain better model performance, the present invention adds the step of supervised learning:

[0060] Step S4: The encoding neural network whose parameters are updated in step S3 is used to extract the feature vector of each sample in the source domain dataset, and the decoding neural network is used to construct a pseudo sample of the target domain based on the feature vector of each sample obtained to obtain a pseudo target domain dataset; the classification loss, the first joint distribution loss, and the comprehensive loss corresponding to the multiple distribution differences between the pseudo target domain dataset and the target domain fully labeled dataset with the pseudo labels obtained in step S3 are used to judge whether the transfer learning meets the predetermined requirements. If the predetermined requirements are not met, the transfer learning is continued.

[0061] According to one embodiment of the present invention, after updating the parameters of the encoding neural network using all samples in the source domain dataset and the target domain dataset each time in step S3, step S4 is executed while locking the parameters of the encoding neural network, and when it is necessary to continue transfer learning, the parameters of the encoding neural network are unlocked to continue transfer learning.

[0062] According to one embodiment of the present invention, step S4 includes: S41, using a decoding neural network to construct a pseudo sample of the target domain based on the feature vector of each sample in the source domain extracted by the encoding neural network whose parameters are updated in step S3, to obtain a pseudo target domain dataset; S42, decomposing the target domain fully annotated dataset and the pseudo target domain dataset into feature sets with different information, and calculating the comprehensive loss corresponding to the difference between the target domain fully annotated dataset and the pseudo target domain dataset and the feature sets corresponding to the decomposition of the target domain fully annotated dataset and the pseudo target domain dataset; S43, judging whether the classification loss, the first joint distribution loss and the comprehensive loss have all converged, if so, ending the transfer learning, if not, continuing to perform the transfer learning.

[0063] According to one embodiment of the present invention, still referring to Figure 3The target domain fully annotated dataset B' and the pseudo target domain dataset AB' are respectively decomposed into the following feature sets: the target domain period feature set BS with period information and the pseudo target domain period feature set AB'S; the target domain trend feature set BT with trend information and the pseudo target domain trend feature set AB'T; the target domain period multiple feature set Fbs with period multiple information and the pseudo target domain period multiple feature set Fab's. According to one embodiment of the present invention, a time series decomposition algorithm (Seasonal-Trend decomposition using LOESS, also known as the STL algorithm; the STL algorithm is a filtering algorithm that can decompose a time series into a trend term, a periodic term (seasonal term), and a residual term. LOESS (locally weighted regression) is a non-parametric method for local regression analysis. It primarily divides samples into small intervals, performs polynomial fitting on the samples within the intervals, and repeats this process to obtain weighted regression curves for different intervals. Finally, the centers of these regression curves are connected to form a complete regression curve) is used to decompose the target domain fully annotated dataset B' and the pseudo-target domain dataset AB'. This yields a target domain periodic feature set BS and a pseudo-target domain periodic feature set AB'S with periodic information, as well as a target domain trend feature set BT and a pseudo-target domain trend feature set AB'T with trend information. Fourier transforms are performed on the target domain periodic feature set BS and the pseudo-target domain periodic feature set AB'S, respectively, to obtain a target domain period multiple feature set Fbs and a pseudo-target domain period multiple feature set Fab's. Correspondingly, the comprehensive loss includes: the second joint distribution loss including marginal distribution difference and conditional distribution difference between the target domain fully annotated dataset B' and the pseudo target domain dataset AB', the third joint distribution loss including marginal distribution difference and conditional distribution difference between the target domain period feature set BS and the pseudo target domain period feature set AB'S, the cosine similarity between the target domain trend feature set BT and the pseudo target domain trend feature set AB'T, and the fourth joint distribution loss including marginal distribution difference and conditional distribution difference between the target domain period multiple feature set Fbs and the pseudo target domain period multiple feature set Fab's. The technical solution of this embodiment can at least achieve the following beneficial technical effects: the present invention judges whether the transfer learning reaches the predetermined requirements based on the comprehensive loss corresponding to the multiple distribution differences between the pseudo target domain dataset constructed by the decoding neural network and the target domain fully annotated dataset, so as to enable the encoding neural network to better learn the knowledge of reducing the distribution difference between the source domain and the target domain, and better improve the accuracy of classification based on the feature vector extracted by the encoding neural network.

[0064] In order to facilitate the explanation of the beneficial effects brought about by step S4, the applicant gives an example here: for periodic signals, their data distribution can be divided into time domain distribution and frequency domain distribution, and the time and frequency domains jointly describe the distribution characteristics of a complete periodic signal. For the time domain waveform of a periodic signal, at least 5 data points are required to describe a complete cycle (sine wave) (at this time, the sine wave is close to a triangle wave and is on the verge of distortion). Taking vibration velocity signal acquisition as an example, the frequency band of 0-1000Hz is valid information, so the sampling frequency is usually 2560 points / second. If only the time domain distribution is observed, the complete periodic information that can be extracted by the convolutional neural network can only reach up to 2560 / 5=512Hz. That is, if only the time domain distribution is considered in the migration of 500-1000Hz, information loss will occur. Therefore, it is necessary to introduce frequency domain distribution migration in the migration process. For the vibration speed signal of rotating dynamic equipment, the equipment speed is usually between 900-3000 rpm, that is, the energy peak frequency band is between 15-50Hz, and after 50Hz, the energy will become an integer multiple of its frequency domain (such as 2 times the frequency is between 30-100, 3 times the frequency is between 45-150, etc.). Therefore, from the perspective of the overall time domain information, the trend information of 0-10Hz will be masked, so it is necessary to extract the trend separately to amplify the distribution characteristics of 0-10Hz. The trend feature extracted in the present invention can well amplify the time-frequency domain distribution characteristics of 0-10Hz, the periodic feature can well retain the time-frequency domain characteristics of 10-500Hz, and the period multiple feature extracted after the frequency domain conversion can amplify the frequency domain energy of 500-1000Hz and convert it into an impact energy signal; thereby achieving a better reduction of the joint distribution (marginal distribution + conditional distribution) than simple time domain migration.

[0065] It should be understood that the previous embodiment is merely a preferred implementation, and the losses included in the comprehensive loss can be set or adjusted based on the datasets decomposed by the implementer during application. In other words, the losses included in the comprehensive loss are related to the feature sets decomposed from the target domain fully annotated dataset B' and the pseudo-target domain dataset AB'. That is, if the implementer decomposes the target domain fully annotated dataset B' and the pseudo-target domain dataset AB' into corresponding feature sets, the comprehensive loss will include the corresponding loss terms. For example, according to one embodiment of the present invention, the target domain fully annotated dataset B' and the pseudo-target domain dataset AB' are decomposed into the following feature sets: a target domain periodic feature set BS with periodicity information and a pseudo-target domain periodic feature set AB'S; a target domain trend feature set BT with trend information and a pseudo-target domain trend feature set AB'T. Correspondingly, the comprehensive loss includes: the second joint distribution loss between the target domain fully annotated dataset B' and the pseudo-target domain dataset AB'; the third joint distribution loss of the distribution difference between the target domain periodic feature set BS and the pseudo-target domain periodic feature set AB'S; and the cosine similarity between the target domain trend feature set BT and the pseudo-target domain trend feature set AB'T. For another example, according to one embodiment of the present invention, the target domain fully annotated dataset B' and the pseudo target domain dataset AB' are respectively decomposed into the following feature sets or their combinations: the target domain trend feature set BT with trend information and the pseudo target domain trend feature set AB'T; the target domain period multiple feature set Fbs with period multiple information and the pseudo target domain period multiple feature set Fab's. Correspondingly, the comprehensive loss includes: the second joint distribution loss between the target domain fully annotated dataset B' and the pseudo target domain dataset AB', the cosine similarity between the target domain trend feature set BT and the pseudo target domain trend feature set AB'T, and the fourth joint distribution loss between the target domain period multiple feature set Fbs and the pseudo target domain period multiple feature set Fab's. It should be understood that there are other possible combinations of the decomposed datasets and the losses contained in the comprehensive loss, which can be set by the implementer as needed and will not be elaborated here.

[0066] According to one embodiment of the present invention, a unified predetermined threshold can be set for determining whether the classification loss, the first joint distribution loss, and the comprehensive loss have all converged. When the classification loss, the first joint distribution loss, and the comprehensive loss are all less than the predetermined threshold for a predetermined number of consecutive times, the classification loss, the first joint distribution loss, and the comprehensive loss are considered to have converged. Alternatively, a separate predetermined threshold can be set for each of the classification loss, the first joint distribution loss, and the comprehensive loss. When each of the classification loss, the first joint distribution loss, and the comprehensive loss is less than the corresponding predetermined threshold for a predetermined number of consecutive times, the classification loss, the first joint distribution loss, and the comprehensive loss are considered to have converged.

[0067] According to one embodiment of the present invention, this embodiment differs from the aforementioned embodiment in that this embodiment only includes the aforementioned steps S1, S2, and S3. For example, steps S1-S3 are repeated multiple times until the first joint distribution loss converges, thereby obtaining a trained encoding neural network and classifier. Compared to the prior art of first training the encoding neural network and classifier with a source domain dataset and then fine-tuning with a target domain dataset, this embodiment does not utilize step S4 to supervise whether the distribution differences between the decomposed datasets containing period information, trend information, period multiple information, etc. are reduced. However, due to the use of a mixed training classifier with labeled data from the source and target domains, and the use of the first joint distribution loss to update the encoding neural network to reduce the marginal distribution differences and conditional distribution differences between the two domains, better results can be achieved compared to the prior art.

[0068] IV. Application

[0069] According to one embodiment of the present invention, a method for classifying device status is provided, comprising: obtaining a one-dimensional periodic signal collected from the device and using it as a sample to be predicted; inputting the sample to be predicted into an encoding neural network and a classifier trained using the aforementioned transfer learning method for classification based on one-dimensional periodic signals, and outputting the device status category. The sample to be predicted is a one-dimensional periodic signal collected from the device. It should be understood that the one-dimensional periodic signal contained in the sample to be predicted can also be collected from the device corresponding to the target domain.

[0070] According to one embodiment of the present invention, a transfer learning system for classifying one-dimensional periodic signals is provided, comprising an encoding neural network and a classifier. The transfer learning system is configured to train the encoding neural network and the classifier according to the aforementioned transfer learning method. Preferably, the transfer learning system may also include the aforementioned decoding neural network. The detailed description of the encoding neural network, classifier, and decoding neural network can be found in the aforementioned embodiments and will not be repeated here.

[0071] In order to verify the effect of the present invention, the applicant conducted an experiment. The experiment collected vibration data on two different types of equipment and used the technology of the present invention to improve the accuracy of mechanical fault classification. The experimental case is as follows: vibration acceleration signals were collected on a SpectraQuest mechanical fault simulator. A total of four acceleration sensors were deployed on the mechanical fault simulator, two of which were installed on the non-drive end bearing seat, with the installation directions being vertical and horizontal respectively; two were installed on the drive end bearing seat, with the installation directions also being vertical and horizontal respectively. The simulator simulated a total of five types of faults, namely: imbalance, misalignment, eccentricity, looseness, and bearing damage. The simulator's fault data and normal operation data were used as the source domain data set for transfer learning. A physical centrifugal split pump was selected as the test equipment, and vertical and horizontal acceleration sensors were installed on the bearing seats at both ends of the pump. The normal operation data of the pump was collected as the target domain data set for transfer learning (that is, in the experiment, only normal samples were marked in the target domain data). A real-world pump failure experiment was conducted, and three types of failures were collected as test data: imbalance, misalignment, and looseness (i.e., in the target domain test data, samples with the failure states of imbalance, misalignment, and looseness were labeled). The actual pump data in the data came from the factory inspection test bench of the pump manufacturer, and the rotor test bench data came from the test bench of the self-built mechanical failure laboratory. The encoding neural network used in transfer learning is Figure 1 The structure shown, the decoding neural network is Figure 2 The structure shown in the figure is shown; in terms of hardware conditions, the data acquisition sensor sampling frequency is 51200, the sampling time is 0.64 seconds, and the computer used is a computer with an NVIDIA GPU. Based on experimental data, the actual test results of the present invention are compared with those of the prior art: 1) According to the prior art, a CNN classifier is trained using source domain training data to form a model, and the model is fine-tuned (Finetune) using target domain training data. The prediction accuracy of the test data is 68%. 2) The present invention uses both source and target domain training data. Since the target domain training data are all labeled normal, the distribution of the normal labeled data in the source domain training data and the target domain data is trained to be as similar as possible. At the same time, the distance between the class of the other labels in the target domain distribution and the normal labels is maintained during training, and the distance between the class of the labels is maintained. The prediction accuracy of the test data is 85%. It can be seen that even if only samples of one state category are annotated in the training target domain dataset, the transfer effect of the present invention is still good. If samples of more state categories are annotated in the target domain dataset, it is expected that the transfer effect of the present invention will be further improved.

[0072] It should be noted that although the above describes the various steps in a specific order, it does not mean that the steps must be performed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order as long as the required functions can be achieved.

[0073] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0074] A computer-readable storage medium may be a tangible device that holds and stores instructions used by an instruction execution device. Computer-readable storage media may include, for example, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove having instructions stored thereon, and any suitable combination thereof.

[0075] While various embodiments of the present invention have been described above, the above descriptions are intended to be illustrative, non-exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or technological improvements in the marketplace, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A transfer learning method for classification based on one-dimensional periodic signals, characterized in that: The method comprises the following steps: S1. Using an encoding neural network to extract feature vectors of each sample in a source domain dataset and a target domain dataset, using the feature vectors and labels of labeled samples in the source domain and the target domain to train a classifier to identify the state category of the device, and updating the parameters of the classifier according to the classification loss, wherein the samples in the source domain and the target domain are collected from different devices, all samples in the source domain are labeled, and some samples in the target domain are labeled, each label indicates the state category corresponding to the device when the corresponding sample is collected, and the sample is a one-dimensional periodic signal collected for the device, and the one-dimensional periodic signal is a vibration signal collected by a sensor; S2. Input the feature vector of each sample without a label in the target domain into the classifier trained in step S1 to output the state category corresponding to the sample, and use it as the pseudo label of the corresponding sample; S3. Determine a first joint distribution loss between the source domain and the target domain, including marginal distribution differences and conditional distribution differences, based on the feature vectors and labels of the samples in the source domain dataset and the feature vectors, labels, and pseudo-labels of the samples in the target domain dataset, and update the parameters of the encoding neural network based on the first joint distribution loss. S4. Use the encoding neural network whose parameters are updated in step S3 to extract the feature vector of each sample in the source domain dataset, and use the decoding neural network to construct a pseudo sample of the target domain based on the obtained feature vector of each sample to obtain a pseudo target domain dataset; judge whether the transfer learning meets the predetermined requirements based on the classification loss, the first joint distribution loss, and the comprehensive loss corresponding to the multiple distribution differences between the pseudo target domain dataset and the target domain fully labeled dataset with the pseudo labels obtained in step S3. If the predetermined requirements are not met, continue to perform transfer learning.

2. The method according to claim 1, characterized in that After updating the parameters of the encoding neural network using all samples in the source domain dataset and the target domain dataset each time in step S3, step S4 is executed while locking the parameters of the encoding neural network. When transfer learning needs to be continued, the parameters of the encoding neural network are unlocked to continue transfer learning.

3. The method according to claim 2, characterized in that The step S4 comprises: S41, using a decoding neural network to construct a pseudo sample of the target domain based on the feature vector of each sample in the source domain extracted by the encoding neural network whose parameters are updated in step S3, to obtain a pseudo target domain dataset; S42, decomposing the target domain fully annotated dataset and the pseudo target domain dataset into feature sets with different information, and calculating a comprehensive loss corresponding to the distribution difference between the target domain fully annotated dataset and the pseudo target domain dataset, and between the feature sets corresponding to the decomposition of the target domain fully annotated dataset and the pseudo target domain dataset; S43. Determine whether the classification loss, the first joint distribution loss, and the comprehensive loss have all converged. If not, continue to perform transfer learning.

4. The method according to claim 3, characterized in that The target domain fully annotated dataset and the pseudo target domain dataset are decomposed into the following feature sets or their combinations: Target domain periodic feature set with periodic information and pseudo target domain periodic feature set; Target domain trend feature set with trend information and pseudo target domain trend feature set; Target domain period multiple feature set with period multiple information and pseudo target domain period multiple feature set.

5. The method according to claim 4, characterized in that The comprehensive loss includes: a second joint distribution loss including marginal distribution difference and conditional distribution difference between the target domain fully annotated dataset and the pseudo target domain dataset, a third joint distribution loss including marginal distribution difference and conditional distribution difference between the target domain period feature set and the pseudo target domain period feature set, a cosine similarity between the target domain trend feature set and the pseudo target domain trend feature set, a fourth joint distribution loss including marginal distribution difference and conditional distribution difference between the target domain period multiple feature set and the pseudo target domain period multiple feature set, or a combination thereof. Among them, the loss included in the comprehensive loss is related to the feature sets decomposed from the fully labeled target domain dataset and the pseudo target domain dataset respectively.

6. The method according to any one of claims 1 to 5, characterized in that The encoding neural network is a down-sampled one-dimensional convolutional neural network, which is used to perform multiple batch normalization, convolution processing and down-sampling processing on the sample data, and then output the feature vector of the sample through global average pooling processing.

7. The method according to claim 6, characterized in that The encoding neural network includes a plurality of convolutional layers for convolution processing, The width of the convolution kernel of the first convolution layer is set according to the number of correlations of the signal source of the input one-dimensional periodic signal, and the length of the convolution kernel of the first convolution layer is the least common multiple of the period lengths of all one-dimensional periodic signals in the source domain dataset and the target domain dataset; and The size of the convolution kernel of the convolution layer after the first convolution layer is selected from a plurality of preset convolution kernel sizes. The size of the convolution kernel of the convolution layer after the first convolution layer is equal to or smaller than the size of the convolution kernel of the previous convolution layer.

8. The method according to claim 6, characterized in that The decoding neural network is an up-sampled one-dimensional convolutional neural network, which is used to perform multiple deconvolution and up-sampling processes on the feature vectors of the samples in the source domain and output pseudo samples in the target domain.

9. A method for classifying device status, characterized in that: include: Obtain a one-dimensional periodic signal collected from the device and use it as a sample to be predicted; The sample to be predicted is input into the encoding neural network and classifier obtained by the method according to any one of claims 1 to 8, and the state category of the device is output.

10. A transfer learning system for classifying one-dimensional periodic signals, comprising: The encoding neural network and classifier are characterized in that the transfer learning system is configured to: train the encoding neural network and classifier according to the method according to any one of claims 1-8.

11. The transfer learning system according to claim 10, wherein: The encoding neural network is a down-sampled one-dimensional convolutional neural network, which is used to perform multiple batch normalization, convolution processing and down-sampling processing on the sample data, and then output the feature vector of the sample through global average pooling processing; The classifier is a fully connected network classifier, and the classifier identifies the status category of the device according to the feature vector of the sample.

12. The transfer learning system according to claim 10, wherein: The transfer learning system also includes a decoding neural network, which is an up-sampled one-dimensional convolutional neural network, configured to perform multiple deconvolution and up-sampling processes on the feature vectors of samples in the source domain and output pseudo samples in the target domain.

13. A computer-readable storage medium, characterized in that A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of the method according to any one of claims 1 to 8 and 9.

14. An electronic device, characterized in that: include: one or more processors; as well as a memory, wherein the memory is used to store executable instructions; The one or more processors are configured to implement the steps of the method of any one of claims 1 to 8 and 9 by executing the executable instructions.

Images