Remote motor real-time monitoring and early warning system based on deep learning

Abstract

The invention discloses a remote motor real-time monitoring and early warning system based on deep learning. The remote motor real-time monitoring and early warning system comprises a magnetic type acceleration sensor, a single-chip microcomputer, an ESP wireless transmission module and an upper computer, the magnetic type acceleration sensor is installed on a motor driving end shell and used for collecting vibration signals in real time and transmitting the collected vibration signals to the single-chip microcomputer to be preprocessed, and the preprocessed signals are sent to the upper computer through the ESP wireless transmission module. A feature extraction module in the upper computer performs multi-scale feature extraction on the vibration signals by adopting a VMD (variational mode decomposition) and FFT (fast Fourier transform) method; the fault detection module combines an improved Swin-Transform model and an LSTM layer to extract spatial features and time sequence features respectively, and generates a final classification result through feature fusion; and the early warning module generates early warning information of different levels according to the fault detection result, and displays the early warning information to monitoring personnel through a front-end page. The method can stably work under variable working conditions, and higher-precision fault prediction and diagnosis are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and particularly relates to a remote motor real-time monitoring and early warning system based on deep learning. Background Art

[0002] Mechanical fault diagnosis is an important means to ensure the stable operation of equipment and reduce maintenance costs. With the advancement of Industry 4.0 and intelligent manufacturing, the real-time monitoring and predictive maintenance of mechanical equipment have become increasingly crucial. However, traditional fault diagnosis methods mainly rely on manual experience and simple signal processing techniques, such as FFT and time-domain statistics, and have limitations when facing complex working conditions and large-scale data. These methods cannot effectively process multi-frequency components or non-linear and non-stationary signals, resulting in a decrease in diagnostic accuracy, especially when the frequency changes.

[0003] Vibration signal analysis is one of the most commonly used techniques in fault diagnosis, and the health status of bearings directly affects the performance of the entire mechanical system. Through effective feature extraction and pattern recognition, early warning of bearing faults can be achieved. However, existing vibration signal analysis methods, such as FFT and WT, have the following problems:

[0004] 1. Insufficient multi-scale feature extraction: Traditional methods usually can only extract features of a single scale and cannot comprehensively capture the details of multi-frequency signals. For example, FFT cannot process signals with changing frequencies and has obvious limitations.

[0005] 2. Insufficient processing of time dependence: Existing methods usually ignore the time series relationship in the signal and are difficult to capture the dynamic changes in long time series, resulting in poor performance when processing signals with strong time dependence.

[0006] 3. Inefficient manual feature engineering: Traditional feature engineering requires a large amount of domain knowledge and cannot be automatically adjusted under different working conditions and fault modes, resulting in low feature extraction efficiency and increased development costs.

[0007] In addition, existing models are prone to overfitting when the data is limited or the working conditions change, resulting in poor generalization ability. Traditional models are designed based on fixed assumptions and cannot cope with real-time data or changes in working conditions, thus unable to provide effective diagnostic results.

[0008] In terms of sensor acquisition, traditional acceleration sensors mostly rely on wired connections, which are complex in wiring and high in maintenance cost, and are difficult to meet the needs of large or remote facilities. Wired transmission may cause signal delay or data loss, affecting the accuracy and real-time performance of fault diagnosis.

[0009] With the development of deep learning technology, CNN and RNN (especially LSTM and GRU) have been gradually applied to mechanical fault diagnosis due to their powerful feature extraction capabilities and excellent processing capabilities for time series. However, existing deep learning models still face challenges in multi-scale feature fusion and time-dependence processing:

[0010] 1. The problem of multi-scale feature fusion: The features at different scales vary greatly. How to fuse these features without losing important information is a key issue. For example, low-scale features capture detailed information, while high-scale features reveal global trends. Traditional models often struggle to balance the two, resulting in feature loss or over-simplification.

[0011] 2. The difference in physical meaning of scales: Low-frequency signals and high-frequency signals have different physical backgrounds. Traditional methods are difficult to ensure that they do not interfere with each other during fusion, leading to a decrease in diagnostic accuracy.

[0012] 3. Poor scale adaptability: In multi-scale feature fusion, traditional methods fail to dynamically adjust the weights of features at different scales, resulting in the model being unable to adapt to feature changes under different working conditions, thereby affecting the generalization ability and diagnostic accuracy of the model. Summary of the Invention

[0013] To address the above technical problems, in order to fully exploit the fault information in different-scale time domains and their transform domains and utilize the characteristics of different deep learning models, this technical solution provides a remote motor real-time monitoring and early warning system based on deep learning. The information processing method is based on multi-domain information fusion, combining frequency domain and time domain for analysis. First, variational mode decomposition (VMD) is used to decompose the vibration signal into multiple intrinsic mode function (IMF) components, and then fast Fourier transform (FFT) is performed. Then, the IMF components and their corresponding frequency domain information are input into an improved deep learning model. The improved deep learning model uses Swin-Transformer and LSTM for modeling. Through multi-scale feature extraction and time series analysis, combined with feature fusion and regularization techniques, it effectively solves the deficiencies of existing methods in multi-scale information capture and time-dependence processing, and effectively solves the above problems.

[0014] The present invention is achieved through the following technical solutions:

[0015] A remote motor real-time monitoring and early warning system based on deep learning, comprising:

[0016] Magnetic adsorption acceleration sensor: Installed on the outer shell of the motor drive end to collect the vibration signal of the motor in real time and transmit the collected vibration signal to the single-chip microcomputer;

[0017] Single-chip microcomputer: Connected to the magnetic adsorption acceleration sensor to receive and preprocess the vibration signal;

[0018] ESP wireless transmission module: Connected to the single-chip microcomputer, it wirelessly transmits the vibration signal preprocessed by the single-chip microcomputer to the host computer;

[0019] Host computer: Includes an ESP communication module, a feature extraction module, a fault detection module, and a warning module; The fault detection module performs fault detection on the extracted features based on an improved Swin-Transformer model to generate a fault detection result;

[0020] The described ESP communication module receives the vibration signal from the ESP wireless transmission module; The feature extraction module uses the methods of VMD variational mode decomposition and FFT fast Fourier transform to perform multi-scale feature extraction on the vibration signal; The fault detection module combines the Swin-Transformer model and the LSTM layer to form a multi-branch structure, extracts spatial features and time series features respectively, and generates a final classification result through feature fusion; The warning module generates warning information at different levels according to the fault detection result and displays it to the monitoring personnel through the front-end page.

[0021] Furthermore, in the methods of VMD variational mode decomposition and FFT fast Fourier transform, VMD variational mode decomposition is configured to decompose the time-domain frequency components of the signal; The FFT fast Fourier transform is configured to extract the spectral information of the signal.

[0022] Furthermore, in the methods of VMD variational mode decomposition and FFT fast Fourier transform, for the collected vibration signal, first use variational mode decomposition (VMD) to extract multi-scale features in the signal; Specifically:

[0023] Input the original vibration signal into the VMD algorithm, and VMD realizes the adaptive decomposition of the signal by minimizing the following constrained variational problem:

[0024]

[0025] where K is the number of decomposed modes; is the partial derivative symbol, representing the partial derivative with respect to the time variable t; δ(t) is the Dirac function; t is the time variable; j is the imaginary unit; * represents the convolution operation; u k (t) represents the k-th intrinsic mode function (IMF); e is the base of the natural logarithm; ω k is the corresponding central frequency;

[0026] By introducing the Lagrange multiplier λ(t), the following constraint conditions are obtained for optimization:

[0027]

[0028] Solve using the Alternating Direction Method of Multipliers (ADMM), alternately optimizing two variables: the modal function u k (t) and the center frequency ω k , and use the Lagrange multiplier λ(t) for constraint update; finally, VMD decomposes the original vibration signal into multiple IMFs, each IMF representing an inherent oscillation mode of the signal, with different center frequencies and bandwidths; these IMFs can better represent different frequency components in the original signal, thus providing richer information for subsequent feature extraction and classification.

[0029] Furthermore, after obtaining the IMFs decomposed by the VMD, perform a Fast Fourier Transform (FFT) on each IMF to transform it into the frequency domain and extract spectral features; for each IMF U k [n], its Discrete Fourier Transform (DFT) is defined as:

[0030]

[0031] where, U k [m] is the frequency-domain representation of IMF U k [n], N is the length of the IMF, e -j2πmn / N is the complex exponential term, representing the basis function of the discrete Fourier transform, j is the imaginary unit, and 2πmn / N represents the frequency component of the odd function;

[0032] Calculate the spectrogram of each IMF through FFT, and analyze the energy distribution of the signal at different frequencies. Extract features such as peak frequency, bandwidth, and energy ratio from the spectrogram. These spectral features, together with the mean, variance, skewness, and kurtosis in the statistical features of the IMFs, form a comprehensive feature vector, which is then input into the fault detection module for classification and detection tasks.

[0033] Furthermore, the fault detection module includes an input layer, a backbone network, a neck network, feature fusion, a regularization layer, and a head network; the input layer receives vibration signal data; the backbone network is equipped with a Swin-Transformer model for feature extraction to extract the spatial features of the signal; the neck network is equipped with LSTM time series analysis to capture the time dependence of the signal; the feature fusion layer combines the spatial and time signal features of the signal;

[0034] And / or, the implementation effectively fuses spatial features and temporal features. The specific operation method is as follows: Add a linear transformation layer after SwinTransformer and match the output dimension of Swin Transformer with the output dimension of LSTM; then the spatial features extracted by Swin Transformer and the temporal features extracted by LSTM can be effectively fused in the subsequent classification layer; ensure that the spatial features of Swin Transformer and the temporal features of LSTM can be seamlessly connected, improving the model's ability to jointly model spatio-temporal features; and through adaptive pooling operations, further enhance the output of the swin transformer.

[0035] Furthermore, the input layer receives the original vibration signal data and performs necessary preprocessing for subsequent feature extraction. The specific implementation steps are as follows: First, the input layer receives the original vibration signal data with the shape of [batch_size, input_dim, seq_length], where batch_size represents the batch size, input_dim represents the feature dimension at each time point, and seq_length represents the time series length. To make the data meet the input requirements of Swin-Transformer, the input layer reshapes the original vibration signal data into the form of [batch_size, input_dim, 32, 32]. This reshaping operation is achieved by converting the one-dimensional time series data into a two-dimensional spatial structure, so that Swin-Transformer can effectively extract spatial features.

[0036] Furthermore, the Swin-Transformer feature extraction part in the backbone network processes the reshaped input data to facilitate the extraction of spatial features. The specific operation method for the Swin-Transformer model in the backbone network to extract spatial features is as follows: First, process the reshaped input data, and send the input data x with the shape of [batch_size, input_dim, 32, 32] after processing into Swin-Transformer to obtain the output feature map F with the shape of [batch_size, 8, 8, 64]; then, perform an adaptive average pooling operation on the feature map F to compress it to the shape of [batch_size, 64, 1, 1]; finally, flatten the pooled feature map F into a one-dimensional vector with the shape of [batch_size, 64].

[0037] Furthermore, the specific implementation steps of introducing the LSTM algorithm into the neck network are as follows: First, the original input data with the shape of [batch_size, input_dim, seq_length] is directly fed into the LSTM layer. The LSTM layer processes the entire time series data and outputs a hidden state sequence. The hidden state dimension of the LSTM layer is set to lstm_hidden_dim, and the number of layers is set to lstm_num_layers. Through its internal memory units and gating mechanisms (input gate, forget gate, and output gate), the LSTM layer can effectively capture dependencies within a long time range. The output of the LSTM layer is the hidden state of the last time step, with the shape of [batch_size, lstm_hidden_dim]. This output contains important feature information in the time series and can be used for subsequent feature fusion and classification tasks. In this way, the LSTM layer can extract temporal dynamic change features from the vibration signal, thereby improving the classification and detection performance of the model.

[0038] Furthermore, the feature fusion, regularization layer, and head network generate the final classification result; the specific operation method is as follows: In the feature fusion layer, the spatial feature swin_out with the shape of [batch_size, 64] extracted by Swin-Transformer and the temporal feature lstm_out with the shape of [batch_size, lstm_hidden_dim] extracted by LSTM are concatenated along the feature dimension to obtain a multi-scale feature vector combined_features that combines temporal and spatial information, with the shape of [batch_size, 64 + lstm_hidden_dim]; then, in the regularization layer, the concatenated feature vector combined_features is applied to the Dropout layer, randomly deactivating some neurons with a retention probability of p, usually set to 0.5, to prevent overfitting; the feature vector after Dropout processing is denoted as dropout_features; finally, in the head network layer, dropout_features is input into the fully connected layer for linear transformation to generate the final classification result, with the shape of [batch_size, output_dim]; the formula for the fully connected layer is as follows:

[0039] outputs = W · dropout_features + b

[0040] The weight matrix W and bias vector b of the fully connected layer are optimized through the training process to ensure that the model can accurately classify bearing vibration signals.

[0041] Further, when preprocessing the vibration signals of the faulty motor collected by the magnetic adsorption acceleration sensor, they are classified according to different fault types, and then divided into a training set, a validation set, and a test set according to the ratio of 7:2:1. The data is passed through the fault detection module to generate a fault classification model, which is used to identify and classify the vibration signals of the sensors subsequently installed on the motor.

[0042] Further, the magnetic adsorption acceleration sensor has a magnetic adsorption base, which is directly adsorbed on the motor housing at the driving end of the motor and is connected to the single-chip microcomputer. During the operation of the motor, the real-time vibration signals generated by the motor are transmitted into the single-chip microcomputer.

[0043] And / or, the single-chip microcomputer is connected to a power supply and is connected to the ESP wireless transmission module and the magnetic adsorption acceleration sensor through a serial cable. The single-chip microcomputer and the ESP wireless transmission module are integrated and placed at the edge of the motor. The single-chip microcomputer is used to preprocess the vibration data of the acceleration sensor and upload the preprocessed vibration signals to the host computer through the ESP wireless transmission module.

[0044] Further, the warning module of the host computer includes a warning generation unit for generating warning information according to the fault detection results. The warning generation unit generates warning information according to the fault type and fault probability in the fault detection results, and sets the warning levels of low-level warning, medium-level warning, and high-level warning according to the warning information. Then, the warning information is sent to the front-end page, and the front-end page displays the warning information and the real-time state of the motor, including the vibration signal waveform and characteristic values.

[0045] And / or, the front-end page also displays historical fault data and warning information. The historical fault data includes the fault type, fault time, and fault probability. The warning information includes the warning level, warning time, and fault type.

[0046] Further, the host computer also includes a data storage module, which uses a database to store the received vibration signals and fault detection results.

[0047] Beneficial Effects

[0048] A remote motor real-time monitoring and warning system based on deep learning proposed by the present invention has the following beneficial effects compared with the prior art:

[0049] (1) The system integration of this technical solution combines an accelerometer and an ESP communication module. By deeply integrating the functions of these two components, it realizes the efficient acquisition of bearing vibration signals and wireless remote monitoring. The accelerometer is responsible for accurately capturing the vibration signals of the bearing, providing real-time and high-resolution data. The ESP communication module, based on the signal acquisition, ensures the rapid transmission of data to the cloud or monitoring center through an optimized wireless communication protocol. This combination not only improves the accuracy of data acquisition but also significantly enhances the real-time performance and remote accessibility of the system. Different from traditional data systems with separate acquisition and transmission, the wireless remote monitoring function of this system, through the close cooperation between the accelerometer and the ESP module, not only ensures low-latency signal transmission but also avoids data loss problems in traditional methods. Through this combination, the system can monitor the states of multiple bearings in real time over a large range, significantly improving the efficiency and reliability of fault diagnosis.

[0050] (2) The feature extraction module of this technical solution combines the VMD variational mode decomposition and the FFT fast Fourier transform method to enhance the fault detection ability of vibration signals through multi-scale analysis. VMD adaptively decomposes the original signal into multiple intrinsic mode functions (IMFs), and each IMF represents different frequency components of the signal, which helps to capture the complex features of vibration signals from different scales. However, when using VMD alone, although it can effectively extract time-domain features, there are certain limitations in capturing frequency-domain information. The FFT converts the IMFs decomposed by VMD into the frequency domain to further refine the extraction of spectral features, making up for the deficiency of VMD in frequency-domain feature extraction. The combination of VMD variational mode decomposition and FFT fast Fourier transform has the advantage that VMD can adaptively extract multi-scale signal features at the time-domain level, while FFT supplements information from the frequency-domain perspective, forming a complementary effect. When using VMD alone, it may be difficult to comprehensively capture the detailed changes in the frequency domain; when using FFT alone, the time-domain features of the signal may be ignored. By combining VMD and FFT, it is possible to deeply analyze the signal from both the time-domain and frequency-domain dimensions, thereby improving the accuracy and robustness of fault detection and overcoming the limitations of single methods.

[0051] (3) The fault detection module of this system combines an improved Swin-Transformer model with an LSTM layer to fully utilize their advantages in spatial feature extraction and time series modeling. As a local window-based self-attention mechanism model, Swin-Transformer can effectively extract the spatial features of signals and capture complex local patterns. However, Swin-Transformer has certain limitations in dealing with long-term time dependencies. The introduced LSTM layer captures the long-term dependencies in the signals through its unique gating mechanism, making up for the deficiencies of Swin-Transformer in time series modeling. In the combination of the two, Swin-Transformer can extract the local spatial features of signals within a short time window, while LSTM can effectively capture the long-term dependencies of time series, especially in the long-term changes and trends of signals. Combining Swin-Transformer and LSTM enables this system to perform deep feature learning simultaneously in both spatial and temporal dimensions, thus achieving more accurate and robust fault detection and overcoming the limitations of a single model. Description of the Drawings

[0052] Figure 1 It is a schematic diagram of the overall system architecture of the present invention.

[0053] Figure 2 It is a schematic diagram of the overall process of the present invention.

[0054] Figure 3 It is a flowchart of the algorithm inside the host computer in the present invention.

[0055] Figure 4 It is an algorithm architecture diagram inside the host computer in the present invention.

[0056] Figure 5 It is a framework diagram of the host computer in the present invention.

[0057] Figure 6 It is a t-sne diagram of classification before and after adopting the VMD+FFT feature extraction method in the present invention.

[0058] Figure 7 It is an accuracy-loss curve graph of model training in the present invention.

[0059] Figure 8 It is a classification confusion matrix graph of the model in the test set in the present invention.

[0060] Reference numerals in the drawings: 1 - motor, 2 - magnetic adsorption acceleration sensor, 3 - single-chip microcomputer, 4 - ESP wireless communication module, 5 - host computer. Detailed Embodiments

[0061] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Without departing from the design concept of the present invention, various modifications and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope of the present invention.

[0062] Embodiment 1:

[0063] As Figure 1 shown, a real-time remote motor monitoring and warning system based on deep learning includes a magnetic acceleration sensor 2, a single-chip microcomputer 3, an ESP wireless communication module 4, and a host computer 5. The host computer 5 is equipped with an ESP wireless communication module, a feature extraction module, a fault detection module, a warning module, a front-end page, and a storage module.

[0064] Magnetic acceleration sensor: It is provided with a magnetic base, which is directly adsorbed on the motor housing at the driving end of the motor and is connected to the single-chip microcomputer; it real-time collects the vibration signal of the motor and transmits the collected vibration signal to the single-chip microcomputer. The magnetic acceleration sensor is small in size and light in weight, has the characteristics of convenient installation without the need to disassemble the motor, and will not affect the normal operation of the motor; the sensor is magnetically adsorbed on the motor housing at the driving end of the motor and is connected to the single-chip microcomputer. During the operation of the motor, it real-time collects the vibration signal generated by the motor and transmits the vibration signal generated by the motor to the single-chip microcomputer.

[0065] Single-chip microcomputer: It is connected to the magnetic acceleration sensor and is used to receive and preprocess the vibration signal; the single-chip microcomputer is connected to a power supply and is connected to the ESP wireless transmission module and the magnetic acceleration sensor through a serial cable; the single-chip microcomputer and the ESP wireless transmission module are an integral whole and are placed at the edge of the motor; the single-chip microcomputer is used to preprocess the vibration data of the acceleration sensor and upload the preprocessed vibration signal to the host computer through the ESP wireless transmission module.

[0066] ESP wireless transmission module: It is connected to the single-chip microcomputer and is used to wirelessly transmit the vibration signal preprocessed by the single-chip microcomputer to the host computer; the ESP wireless transmission module is connected to the single-chip microcomputer through a serial cable and is used to transmit the vibration signal collected by the magnetic acceleration sensor to the host computer.

[0067] Host computer: It includes an ESP communication module, a feature extraction module, a fault detection module, and a warning module; as Figure 5 shown. The fault detection module performs fault detection on the extracted features based on an improved Swin-Transformer model and generates a fault detection result.

[0068] The ESP communication module in the host computer is used to receive vibration signals from the ESP wireless transmission module; the feature extraction module uses the methods of VMD variational mode decomposition and FFT fast Fourier transform to perform multi-scale feature extraction on the vibration signals; the fault detection module combines the Swin-Transformer model and the LSTM layer to form a multi-branch structure, extracts spatial features and time series features respectively, and generates the final classification result through feature fusion; the warning module generates warning information at different levels according to the fault detection result and displays it to the monitoring personnel through the front-end page.

[0069] The warning module of the host computer includes a warning generation unit for generating warning information according to the fault detection result. The warning generation unit generates warning information according to the fault type and fault probability in the fault detection result, and sets the warning levels of low-level warning, medium-level warning and high-level warning according to the warning information; then sends the warning information to the front-end page, and the front-end page displays the warning information and the real-time state of the motor, including the vibration signal waveform and characteristic values; and / or, the front-end page also displays historical fault data and warning information; the historical fault data includes fault type, fault time, fault probability; the warning information includes warning level, warning time, fault type.

[0070] The data storage module in the host computer uses a database to store the received vibration signals and fault detection results. Using a database to store vibration signals and fault detection results facilitates subsequent analysis and query, supports data backup and recovery, and ensures data security.

[0071] The system process is as Figure 2 shown. The specific operation steps are as follows: After the system is started, the magnetic acceleration sensor deployed on the motor first collects vibration signals in real time. The sensor can accurately capture the vibration information during the operation of the motor and transmit data wirelessly. The acceleration sensor collects data at a high sampling rate, and collects the vibration frequency of the motor in real time to ensure that enough details are captured.

[0072] The collected vibration signals are transmitted to the single-chip microcomputer, which is responsible for preliminary processing and packaging of the data to ensure that the data format is correct. Then, the single-chip microcomputer sends the data to the host computer through the ESP wireless transmission module. The ESP wireless transmission module ensures wireless transmission of the data, enables the system to be flexibly deployed in a complex industrial environment, reduces wiring requirements and maintenance costs.

[0073] After the host computer receives the data, it immediately performs preprocessing. The preprocessing steps include denoising and normalization operations to improve the data quality. The denoising method uses wavelet transform or filters to ensure the removal of unnecessary noise components. Normalization ensures that the data collected by different sensors have the same dimension and range, facilitating subsequent analysis. The preprocessed vibration signal enters the feature extraction module, which uses a method combining VMD (Variational Mode Decomposition) and FFT (Fast Fourier Transform).

[0074] The feature vector is then sent to the fault detection module, which combines the Swin-Transformer and LSTM models for modeling. According to the classification results generated by the fault detection module, the system will trigger corresponding warning signals. If a fault type is detected, the system will display specific fault information and recommended measures on the front-end page to help engineers discover potential problems in a timely manner. The front-end page not only shows the fault type but also provides a historical data chart function so that users can comprehensively understand the health status of the device.

[0075] As Figure 3 and Figure 4 shown, the algorithm flow inside the host computer is as follows:

[0076] First, the collected fault vibration signals are classified according to the fault type and stored in a local or cloud database. Using the labeled fault data set, it is divided into a training set, a validation set, and a test set according to the ratio of 7:2:1.

[0077] After the host computer receives the vibration signal, it first performs preprocessing operations, including denoising and normalization, to improve the data quality.

[0078] The preprocessed vibration signal enters the feature extraction module, which uses a method combining VMD (Variational Mode Decomposition) and FFT (Fast Fourier Transform). The specific steps are as follows: Step 1: Multi-scale decomposition

[0079] First, use VMD to perform multi-scale decomposition on the vibration signal, decomposing the original signal into multiple IMFs, and each IMF represents an inherent oscillation mode of the signal. VMD realizes the adaptive decomposition of the signal by minimizing the following constrained variational problem:

[0080]

[0081] where is the partial derivative symbol, representing the partial derivative with respect to the time variable t; K is the number of decomposed modes, δ(t) is the Dirac function, t is the time variable, j is the imaginary unit, * represents the convolution operation, u k (t) represents the k-th intrinsic mode function (IMF), e is the base of the natural logarithm, ω kis the corresponding center frequency.

[0082] Finally, VMD decomposes the original vibration signal into multiple IMFs. Each IMF represents an inherent oscillation mode of the signal, with different center frequencies and bandwidths. These IMFs can better represent different frequency components in the original signal, thus providing richer information for subsequent feature extraction and classification.

[0083] Step 2: Frequency domain conversion

[0084] Next, perform FFT transformation on each IMF to convert the time-domain signal into a frequency-domain signal. FFT calculates the discrete Fourier transform (DFT) through the following formula:

[0085]

[0086] where U k [m] is the frequency-domain representation of IMF U k [n], N is the length of the IMF, e -j2πmn / N is the complex exponential term, representing the basis function of the discrete Fourier transform, and j is the imaginary unit. Through FFT, the spectrogram of each IMF can be efficiently calculated, thereby analyzing the energy distribution of the signal at different frequencies; 2πmn / N represents the frequency components of the odd function;

[0087] From the spectrogram, a series of features can be extracted, such as peak frequency, bandwidth, energy ratio, etc. These spectral features, together with the statistical features of the IMFs (such as mean, variance, skewness, kurtosis, etc.), constitute a comprehensive feature vector. As shown in the classification result Figure 6 , in this figure, the more dispersed the distribution of various fault samples, the more significant its classification efficiency. From Figure 6 observation, it can be seen that the initial fault samples are extremely messy in the space and there is a serious problem of sample overlap, which directly affects the quality of the classification result. In contrast, after the features are extracted and processed by the VMD+FFT algorithm, the arrangement of the samples in the low-dimensional space. After this processing, the samples show better aggregation compared to the original data, and the discrimination between classes is also significantly enhanced. This feature vector is then input into the fault detection module for further classification and detection tasks.

[0088] The feature vector is then sent to the fault detection module. This module combines the Swin-Transformer and LSTM models for modeling. After feature extraction, the data extracts spatial features through the Swin-Transformer and captures time dependencies through the LSTM layer. Step 3: Swin-Transformer extracts spatial features

[0089] The input data after feature extraction is fed into the Swin-Transformer, and spatial features are extracted through multiple layers of Transformer blocks. The output feature map of the Swin-Transformer is denoted as, with a shape of [batch_size, 8, 8, 64]. Then, an adaptive average pooling operation is performed on the feature map to compress it to [batch_size, 64], and then flattened into a one-dimensional vector swin_out. The formula for the multi-head self-attention mechanism of the Swin-Transformer is:

[0090] MultiHead(Q, K, V) = Concat(head 1 , …, head h )W O

[0091] where W O is the final linear projection matrix. The dot-product attention calculation formula is:

[0092]

[0093] where d k represents the dimension of the key vector. The Swin-Transformer processes the input two-dimensional feature map (converted from the preprocessed vibration signal) through multiple layers of self-attention mechanisms to extract local and global spatial features layer by layer, and finally obtains a compact feature representation through a pooling operation.

[0094] Step 4: LSTM extracts temporal features

[0095] The original input data with a shape of [batch_size, input_dim, seq_length] is directly fed into the LSTM layer. The LSTM layer processes the entire time series data and outputs a sequence of hidden states. The hidden state dimension of the LSTM layer is set to lstm_hidden_dim, and the number of layers is set to lstm_num_layers. Through its internal memory cells and gating mechanisms (input gate, forget gate, and output gate), the LSTM layer can effectively capture dependencies within a long time range. The output of the LSTM layer is the hidden state of the last time step, with a shape of [batch_size, lstm_hidden_dim]. This output contains important feature information in the time series and can be used for subsequent feature fusion and classification tasks.

[0096] Step 5: Feature fusion, regularization, and fully connected output

[0097] Then, the spatial features swin_out of Swin-Transformer and the temporal features lstm_out of LSTM are combined along the feature dimension to form a comprehensive feature vector combined_features, whose shape is [batch_size, 64 + lstm_hidden_dim]. Next, in the regularization layer, the Dropout layer is applied to the concatenated feature vector combined_features to randomly deactivate some neurons with a retention probability of p (usually set to 0.5) to prevent overfitting. The feature vector after Dropout processing is denoted as dropout_features. Finally, in the head network layer, dropout_features is input into the fully connected layer for linear transformation to generate the final classification result, whose shape is [batch_size, output_dim]. The formula for the fully connected layer is as follows:

[0098] outputs = W · dropout_features + b

[0099] The weight matrix W and bias vector b of the fully connected layer are optimized through the training process to ensure that the model can accurately classify bearing vibration signals.

[0100] Its shape is [batch_size, 64 + lstm_hidden_dim]. Next, in the regularization layer, the Dropout layer is applied to the concatenated feature vector combined_features to randomly deactivate some neurons with a retention probability of p (usually set to 0.5) to prevent overfitting. The feature vector after Dropout processing is denoted as dropout_features. Finally, in the head network layer, dropout_features is input into the fully connected layer for linear transformation to generate the final classification result, whose shape is [batch_size, output_dim]. The formula for the fully connected layer is as follows:

[0101] outputs = W · dropout_features + b

[0102] The weight matrix W and bias vector b of the fully connected layer are optimized through the training process to ensure that the model can accurately classify bearing vibration signals.

[0103] On the training set, the model adjusts its parameters by learning the relationship between the feature vector and the fault type; on the validation set, the model parameters are adjusted to prevent overfitting. The training results are as Figure 7As shown, during training, the loss value of the model gradually decreases as the number of training rounds increases. After approximately 17 rounds of training, the loss value drops below 0.01, demonstrating the high efficiency of the model. At the same time, the classification accuracy also continuously improves as the number of training rounds increases and stabilizes above 99% after approximately 30 rounds, proving the excellent performance of the system proposed in the present invention in this task.

[0104] Figure 8 The presented normalized multi-class confusion matrix further validates the classification ability of the model. The high accuracy (close to 1.0) on the diagonal reflects the excellent discrimination ability of the model for each class. The low values of the off-diagonal elements indicate that the model rarely misclassifies samples into the wrong classes, proving that the method proposed in the present invention has good generalization ability and a low risk of overfitting.

Claims

1. A remote motor real-time monitoring and early warning system based on deep learning, characterized by: include: Magnetic acceleration sensor: installed on the motor drive end housing to collect the motor's vibration signal in real time and transmit the collected vibration signal to the single-chip microcomputer; Single chip microcomputer: connected with magnetic acceleration sensor, receiving and preprocessing vibration signal; ESP wireless transmission module: connected to the single-chip microcomputer, wirelessly transmitting the vibration signal pre-processed by the single-chip microcomputer to the host computer; Host computer: including ESP communication module, feature extraction module, fault detection module and early warning module; the fault detection module performs fault detection on the extracted features based on the improved Swin-Transformer model and generates fault detection results; The ESP communication module receives the vibration signal from the ESP wireless transmission module; the feature extraction module uses VMD variational mode decomposition and FFT fast Fourier transform methods to extract multi-scale features of the vibration signal; the fault detection module combines the Swin-Transformer model and the LSTM layer to form a multi-branch structure, respectively extracts spatial features and time series features, and generates the final classification result through feature fusion; The early warning module generates early warning information of different levels according to the fault detection result, and displays it to the monitoring personnel through the front-end page.

2. A remote motor real-time monitoring and early warning system based on deep learning according to claim 1, characterized in that: In the VMD variational mode decomposition and FFT fast Fourier transform method, the VMD variational mode decomposition is configured to decompose the time domain frequency components of the signal; and the FFT fast Fourier transform is configured to extract the spectrum information of the signal.

3. A remote motor real-time monitoring and early warning system based on deep learning according to claim 1 or 2, characterized in that: In the VMD variational mode decomposition and FFT fast Fourier transform method, for the collected vibration signal, firstly, variational mode decomposition (VMD) is used to extract multi-scale features in the signal; specifically: The original vibration signal is input into the VMD algorithm, and VMD realizes the adaptive decomposition of the signal by minimizing the following constrained variational problem: Where K is the number of decomposed modes, is the symbol of partial derivative, δ(t) is the Dirac function, t is the time variable, j is the imaginary unit, * indicates the convolution operation, u k (t) represents the kth intrinsic mode function (IMF), e is the base of the natural logarithm, ω k is the corresponding center frequency.

4. A remote motor real-time monitoring and early warning system based on deep learning according to claim 3, characterized in that: After obtaining the IMFs after VMD decomposition, perform FFT on each IMF, convert it to the frequency domain and extract the spectrum features; for each IMFU k [n], whose discrete Fourier transform DFT is defined as: Among them, U k [m] is IMFU k [n] frequency domain representation, N is the length of the IMF, e -j2πmn / N is a complex exponential term, representing the basis function of the discrete Fourier transform, j is an imaginary unit, and 2πmn / N represents the frequency component of the odd function; The spectrum of each IMF is calculated by FFT, and the energy distribution of the signal at different frequencies is analyzed to extract the characteristics of peak frequency, bandwidth, and energy ratio from the spectrum. These spectrum features and the mean, variance, skewness, and kurtosis of the statistical characteristics of IMFs constitute a comprehensive feature vector, which is then input into the fault detection module for classification and detection tasks.

5. The remote motor real-time monitoring and early warning system based on deep learning according to claim 1 is characterized in that: The fault detection module includes an input layer, a backbone network, a neck network, feature fusion, a regularization layer, and a head network; the input layer receives vibration signal data; the backbone network is provided with a Swin-Transformer feature extraction model to extract the spatial features of the signal; the neck network is provided with an LSTM time series analysis to capture the temporal dependency of the signal; the feature fusion layer combines the spatial and temporal signal features of the signal; And / or, the effective fusion of spatial features and temporal features is achieved by: adding a linear transformation layer after the Swin-Transformer, and matching the output dimension of the Swin-Transformer with the output dimension of the LSTM; the spatial features extracted by the Swin-Transformer and the temporal features extracted by the LSTM can be effectively fused in the subsequent classification layer; ensuring that the spatial features of the Swin-Transformer and the temporal features of the LSTM can be seamlessly connected, thereby improving the model's ability to jointly model spatial and temporal features; and further enhancing the output of the Swin-Transformer through adaptive pooling operations.

6. A remote motor real-time monitoring and early warning system based on deep learning according to claim 1 or 5, characterized in that: The specific operation method of the improved Swin-Transformer model for extracting spatial features is as follows: first, the reshaped input data is processed, and the input data x with a shape of [batch_size, input_dim, 32, 32] is sent to Swin-Transformer to obtain an output feature map F with a shape of [batch_size, 8, 8, 64]; then, an adaptive average pooling operation is performed on the feature map F to compress it to a shape of [batch_size, 64, 1, 1]; finally, the pooled feature map F is flattened into a one-dimensional vector with a shape of [batch_size, 64].

7. The remote motor real-time monitoring and early warning system based on deep learning according to claim 5 is characterized by: The feature fusion, regularization layer and head network generate the final classification result; the specific operation method is: in the feature fusion layer, the spatial feature swin_out with a shape of [batch_size, 64] extracted by Swin-Transformer and the temporal feature lstm_out with a shape of [batch_size, lstm_hidden_dim] extracted by LSTM are spliced ​​along the feature dimension to obtain a multi-scale feature vector combined_features that integrates temporal and spatial information, and its shape is [batch_size, 64 + lstm_hidden_dim]; then, in the regularization layer, the spliced ​​feature vector combined_features is applied to the Dropout layer, and some neurons are randomly inactivated with a probability of p to prevent overfitting; the feature vector after Dropout processing is recorded as dropout_features; finally, in the head network layer, dropout_features is input into the fully connected layer for linear transformation to generate the final classification result, and its shape is [batch_size, output_dim]; the formula of the fully connected layer is as follows: outputs=W·dropout_features+b The weight matrix W and bias vector b of the fully connected layer are optimized through the training process to ensure that the model can accurately classify the bearing vibration signal.

8. The deep learning-based remote motor real-time monitoring and early warning system according to claim 1, characterized in that: The magnetic acceleration sensor has a magnetic base, which is directly adsorbed on the motor housing at the driving end of the motor and connected to the single-chip microcomputer. When the motor is working, the real-time vibration signal generated by the motor is transmitted to the single-chip microcomputer. And / or, the single-chip microcomputer is connected to a power supply and is connected to the ESP wireless transmission module and the magnetic acceleration sensor through a serial port line; the single-chip microcomputer and the ESP wireless transmission module are integrated as a whole and are placed at the edge of the motor; the single-chip microcomputer is used to pre-process the vibration data of the acceleration sensor and upload the pre-processed vibration signal to the host computer through the ESP wireless transmission module.

9. The remote motor real-time monitoring and early warning system based on deep learning according to claim 1, characterized in that: The warning module of the host computer includes a warning generation unit for generating warning information according to the fault detection result, the warning generation unit generates warning information according to the fault type and fault probability in the fault detection result, and sets the warning level of low-level warning, medium-level warning and high-level warning according to the warning information; Then the warning information is sent to the front-end page, which displays the warning information and the real-time status of the motor, including the vibration signal waveform and characteristic value; And / or, the front-end page also displays historical fault data and warning information; the historical fault data includes fault type, fault time, and fault probability; the warning information includes warning level, warning time, and fault type.

10. The deep learning-based remote motor real-time monitoring and early warning system according to claim 1, characterized in that: The host computer also includes a data storage module, and the data storage module uses a database to store the received vibration signal and the fault detection result.

Images