Spinning machine fault detection method and system based on deep learning

Abstract

The invention relates to the technical field of textile machinery fault detection, and discloses a spinning machine fault detection method and system based on deep learning, and the method comprises the steps: collecting vibration, temperature and current signals through a plurality of sensors, constructing a data stream after principal component analysis and dimension reduction, employing a variational auto-encoder to reconstruct an error positioning abnormal data segment, and obtaining a spinning machine fault detection result. The method comprises the following steps: extracting local features of a multi-source signal in combination with a convolutional neural network, dynamically distributing feature weights by using an attention mechanism, screening time sequence key features through a forgetting gate, introducing an LSTM network to model a feature sequence dynamic relationship, completing fault category mapping based on a support vector machine, and triggering LSTM secondary classification calibration for samples with insufficient confidence. A closed-loop technical system of data dimension reduction, anomaly detection, feature optimization, time sequence modeling and classification verification is formed, and efficient analysis and accurate recognition of multi-modal fault features under complex working conditions are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of spinning machine fault detection, and particularly to a spinning machine fault detection method and system based on deep learning. Background Art

[0002] Currently, the textile industry is an important pillar of the manufacturing industry, and its production efficiency and product quality directly affect the industry's competitiveness. With the advancement of intelligent manufacturing, as the core equipment of textile production, the operating stability and accurate fault detection of spinning machines are crucial for ensuring production continuity and reducing maintenance costs. Traditional fault detection methods mainly rely on manual experience or rule-based monitoring systems. When dealing with complex and variable operating environments of spinning machines, these methods often have the defects of slow response, high false alarm rate, and poor adaptability. Especially for weak abnormal signals in dynamic operation data, existing methods are difficult to accurately identify and classify, resulting in limited accuracy and efficiency of fault diagnosis.

[0003] In an existing technology, first, multi-source time-series data is collected through vibration sensors, temperature probes, and current transformers deployed at key parts of the spinning machine, and the original signal is segmented and intercepted using a sliding window; then, each data segment is preprocessed, high-frequency noise is eliminated through a Butterworth filter, and the dimension is unified using Z-Score standardization; then, time-domain statistical features (such as mean, variance, peak factor) and frequency-domain features (spectrum energy distribution calculated based on FFT) are extracted to form a feature vector with a fixed dimension; for a preset rule base such as vibration amplitude threshold, temperature gradient threshold, etc., a sliding t-test is used to determine whether the feature value exceeds the safe range; for abnormal data segments that trigger the threshold, after dimensionality reduction using principal component analysis, they are input into a support vector machine classifier, and a classification model trained based on historical fault samples will output a preset fault category label, and finally, the detection results and maintenance suggestions are displayed through a human-machine interaction interface.

[0004] In the existing technology, feature engineering based on artificial rules is difficult to capture implicit associations in high-dimensional data, resulting in insufficient discrimination between weak abnormal signals and noise; the fixed threshold system cannot adapt to the dynamic baseline drift during the equipment deterioration process, causing early fault undetected; the single-modal feature extraction strategy breaks the time-series coupling relationship between parameters such as vibration and temperature, weakening the characterization ability of compound faults. The existing technology cannot accurately map to improving the fault detection performance of spinning machines. Summary of the Invention

[0005] The present invention provides a spinning machine fault detection method and system based on deep learning to improve the fault detection performance of spinning machines.

[0006] In a first aspect, to solve the above technical problems, the present invention provides a spinning machine fault detection method based on deep learning, including: Obtain high-dimensional heterogeneous data streams from a spinning machine, including vibration signals, temperature signals, and current signals. Use principal component analysis to retain the main eigenvectors and obtain the first data stream after dimensionality reduction. According to the first data stream, calculate the reconstruction error through the memory cell state and the hidden state. If the reconstruction error exceeds a preset reconstruction error threshold, determine the data segment with the reconstruction error as abnormal to obtain a set of abnormal data segments. Extract the vibration signal, the temperature signal, and the current signal from the set of abnormal data segments, and use a convolutional neural network to extract local features to obtain a first feature set. Use an attention mechanism to perform weighted calculation on the first feature set to obtain a weighted feature set. According to the weighted feature set, use a long short-term memory network to extract a second feature set. According to the second feature set, use a long short-term memory network to analyze the temporal dynamic relationship of the feature sequence to obtain a third feature set containing temporal information. According to the third feature set, use a support vector machine model to map the features to predefined fault categories to obtain a preliminary fault classification result. Obtain the classification confidence from the preliminary fault classification result. If the confidence is lower than a preset confidence threshold, use a long short-term memory network to perform secondary classification on the low-confidence samples to obtain the final fault classification result.

[0007] In an alternative embodiment, the step of calculating the reconstruction error through the memory cell state and the hidden state according to the first data stream, and if the reconstruction error exceeds a preset reconstruction error threshold, determining the data segment with the reconstruction error as abnormal to obtain a set of abnormal data segments includes: Obtain continuous data segments from the first data stream, and divide the continuous data segments by a sliding window method to obtain divided data segments. Use a variational autoencoder to encode and decode the divided data segments, and calculate the reconstruction error through the memory cell and the hidden state to obtain a set of reconstruction errors. If the reconstruction error of any data segment in the set of reconstruction errors exceeds a preset reconstruction error threshold, determine that data segment as abnormal to obtain a set of abnormal data segments.

[0008] In an alternative embodiment, the step of extracting the vibration signal, the temperature signal, and the current signal from the set of abnormal data segments, and using a convolutional neural network to extract local features to obtain a first feature set includes: Extract the vibration signal, the temperature signal, and the current signal from the set of abnormal data segments, and perform denoising and normalization processing to obtain a first signal set. According to the first signal set, the signal sequence is divided by a time window method, and a convolutional neural network is used to extract local features of the signal sequence to obtain a first feature set.

[0009] In an optional implementation, the using of the attention mechanism to perform weighted calculation on the first feature set to obtain a weighted feature set includes: Using an attention mechanism to calculate weight coefficients of the vibration signal, the temperature signal, and the current signal; A weighted feature set is obtained by performing weighted processing according to the first feature set and the weight coefficient.

[0010] In an optional implementation, extracting a second feature set using a long short-term memory network based on the weighted feature set includes: The long short-term memory network calculates a forget gate activation value of the weighted feature set through a forget gate; The forget gate activation value is compared with a preset activation threshold. If the forget gate activation value exceeds the activation threshold, the weighted features of the forget gate activation value are eliminated to obtain a second feature set.

[0011] In an optional implementation, the second feature set is used to analyze the temporal dynamic relationship of the feature sequence using a long short-term memory network to obtain a third feature set containing temporal information, including: Extracting signal sequences of vibration signals, temperature signals and current signals from the second feature set; The signal sequence is divided into time series segments by using a sliding time window method, and time series variation features are extracted; Calculating the weight value of the time series segment through the input gate of the long short-term memory network, and performing weighted processing on the time series change feature to obtain a weighted time series feature; If the weight value of the time sequence segment exceeds a preset weight threshold, the key time sequence segments in the time sequence segments are screened out through the output gate of the long short-term memory network to generate a refined feature set; A dynamic time warping algorithm is used to calculate the morphological similarity of each time sequence segment in the refined feature set, and a feature vector dimension of the refined feature set is adjusted through a gated recurrent unit to obtain a dynamic feature set; According to the dynamic feature set, a decision tree algorithm is used to output a third feature set including fault category labels and timing information.

[0012] In an optional implementation, the classification confidence is obtained from the preliminary fault classification result. If the confidence is lower than a preset confidence threshold, a long short-term memory network is used to perform secondary classification on the low-confidence samples to obtain a final fault classification result, including: According to the preliminary fault classification result, a softmax function is used to calculate the classification confidence of each fault category; If the highest confidence in the classification confidence is lower than the confidence threshold determined in advance through cross-validation, the corresponding third feature is marked as a low-confidence sample; According to the low-confidence samples, vibration signals and temperature signals are extracted from the time-series information of the third feature set, and a long short-term memory network is used to model the feature sequences of the vibration signals and the temperature signals to obtain enhanced feature sequences; Through the enhanced feature sequences, a deep feature fusion classifier is used to perform secondary classification on the low-confidence samples, and a fault classification result after probability calibration is output.

[0013] In a second aspect, the present invention provides a spinning machine fault detection system based on deep learning, including: A data acquisition module, configured to obtain a high-dimensional heterogeneous data stream from a spinning machine, including vibration signals, temperature signals, and current signals, and use principal component analysis to retain the main feature vectors to obtain a first data stream after dimensionality reduction; An abnormal data analysis module, configured to calculate a reconstruction error according to the first data stream through a memory cell state and a hidden state, and if the reconstruction error exceeds a preset reconstruction error threshold, determine that the data segment of the reconstruction error is abnormal to obtain a set of abnormal data segments; A first feature analysis module, configured to extract the vibration signals, the temperature signals, and the current signals from the set of abnormal data segments, and use a convolutional neural network to extract local features to obtain a first feature set; A weighted feature analysis module, configured to perform weighted calculation on the first feature set by using an attention mechanism to obtain a weighted feature set; A second feature analysis module, configured to extract a second feature set according to the weighted feature set by using a long short-term memory network; A third feature analysis module, configured to analyze the temporal dynamic relationship of the feature sequences according to the second feature set by using a long short-term memory network to obtain a third feature set including temporal information; A preliminary fault classification module, configured to map features to predefined fault categories by using a support vector machine model according to the third feature set to obtain a preliminary fault classification result; A final fault classification module, configured to obtain classification confidence from the preliminary fault classification result, and if the confidence is lower than a preset confidence threshold, perform secondary classification on the low-confidence samples by using a long short-term memory network to obtain a final fault classification result.

[0014] In a third aspect, the present invention further provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the method for detecting faults of a spinning machine based on deep learning described in any one of the above is implemented.

[0015] In a fourth aspect, the present invention further provides a computer-readable storage medium, which includes a stored computer program. When the computer program runs, the device where the computer-readable storage medium is located is controlled to execute the method for detecting faults of a spinning machine based on deep learning described in any one of the above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) By constructing a multi-modal data dimensionality reduction framework through variational autoencoders and principal component analysis, the high-dimensional heterogeneous vibration, temperature, and current signals are subjected to feature space mapping and noise filtering, solving the technical bottleneck of high feature redundancy and difficulty in capturing weak anomalies in traditional single-modal analysis methods under complex working conditions, and significantly improving the sensitivity and robustness of abnormal data segment positioning.

[0017] (2) Based on the attention mechanism and forget gate, the temporal correlation of multi-source signal features is modeled. Through dynamic allocation of weight coefficients and iterative elimination of redundant features, the information loss limitation of traditional static feature fusion methods is broken through, enabling accurate extraction of key fault features, effectively reducing the computational complexity while improving the feature representation ability.

[0018] (3) Adopting a cascaded classification architecture of an LSTM network and a support vector machine, the temporal evolution law of fault features is analyzed through the input gate / output gate mechanism, and the non-linear mapping from the fault mode to the preset category is completed in combination with the gradient disappearance mitigation strategy, solving the problem of insufficient generalization ability of traditional rule engines in complex fault mode mapping, and significantly improving the recognition accuracy of multi-type concurrent faults.

[0019] (4) Introducing a confidence-driven secondary classification mechanism, low-confidence samples are screened through soft thresholds and the feature space distribution is reconstructed using LSTM temporal modeling, breaking through the limitations of a single classification model in handling boundary samples, forming a closed-loop diagnosis logic, and greatly improving the decision reliability of the system in scenarios of noise interference and sample imbalance.

[0020] In summary, the present invention constructs a fault detection system for spinning machines that integrates data dimensionality reduction, dynamic feature fusion, time series modeling, and confidence verification. First, it breaks through the dimensionality limitation of traditional methods through multi-modal data collaborative processing; secondly, it uses the attention mechanism and gating unit to achieve dynamic feature optimization; then, it adopts a cascaded classification model to enhance the ability to analyze complex faults; finally, it ensures the system stability through a closed-loop verification mechanism. This technology systematically solves the core problems of traditional detection methods such as response lag, high false alarm rate, and poor generalization, providing a high-precision solution for the intelligent operation and maintenance of textile equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a schematic flowchart of a fault detection method for a spinning machine based on deep learning provided by the first embodiment of the present invention; Figure 2 is a schematic structural diagram of a fault detection system for a spinning machine based on deep learning provided by the second embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] 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. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0023] Refer to Figure 1 , the first embodiment of the present invention provides a fault detection method for a spinning machine based on deep learning, including the following steps: S11, obtaining a high-dimensional heterogeneous data stream from the spinning machine, including vibration signals, temperature signals, and current signals, and retaining the main feature vectors by principal component analysis to obtain a first data stream after dimensionality reduction; S12, calculating the reconstruction error based on the first data stream through the memory cell state and the hidden state. If the reconstruction error exceeds a preset reconstruction error threshold, determining the data segment of the reconstruction error as abnormal to obtain a set of abnormal data segments; S13, extracting the vibration signal, the temperature signal, and the current signal from the set of abnormal data segments, and extracting local features by using a convolutional neural network to obtain a first feature set; S14, performing weighted calculation on the first feature set by using the attention mechanism to obtain a weighted feature set; S15, extracting a second feature set by using a long short-term memory network according to the weighted feature set; S16, analyzing the temporal dynamic relationship of the feature sequence by using a long short-term memory network according to the second feature set to obtain a third feature set including temporal information; S17. According to the third feature set, use a support vector machine model to map the features to predefined fault categories to obtain a preliminary fault classification result; S18. Obtain the classification confidence from the preliminary fault classification result. If the confidence is lower than a preset confidence threshold, use a long short-term memory network to perform secondary classification on the low-confidence samples to obtain the final fault classification result.

[0024] In step S11, obtain a high-dimensional heterogeneous data stream from the spinning machine, including vibration signals, temperature signals, and current signals, and use principal component analysis to retain the main feature vectors to obtain a first data stream with reduced dimensions.

[0025] Specifically, during the operation of the spinning machine, integrate vibration sensors, temperature sensors, and current sensors to collect device status information in real time. The vibration signal captures the micron-level displacement fluctuations of mechanical components through a piezoelectric accelerometer. The temperature signal monitors the thermal distribution gradient on the surface of the bearing and the motor using a thermocouple. The current signal obtains the dynamic electrical parameters of the drive system through a Hall sensor. The high-dimensional heterogeneous data stream is subjected to baseline correction and dimension unification through a signal conditioning module, and an initial high-dimensional data stream is formed after eliminating the influence of environmental noise and sensor drift. The principal component analysis method analyzes the correlation strength between signals by calculating the covariance matrix, extracts the orthogonal basis vectors with the largest eigenvalues to construct a low-dimensional projection space, maps the original data to a feature subset that retains the main energy distribution, realizes data dimension compression and information redundancy elimination, and obtains the first data stream.

[0026] In step S12, according to the first data stream, calculate the reconstruction error through the memory cell state and the hidden state. If the reconstruction error exceeds a preset reconstruction error threshold, determine the data segment with the reconstruction error as abnormal to obtain a set of abnormal data segments.

[0027] In a specific implementation manner, the calculating the reconstruction error through the memory cell state and the hidden state according to the first data stream, and if the reconstruction error exceeds a preset reconstruction error threshold, determining the data segment with the reconstruction error as abnormal to obtain a set of abnormal data segments includes: Obtain continuous data segments from the first data stream, and divide the continuous data segments through a sliding window method to obtain divided data segments; Use a variational autoencoder to encode and decode the divided data segments, and calculate the reconstruction error through the memory cell and the hidden state to obtain a set of reconstruction errors; If the reconstruction error of any data segment in the set of reconstruction errors exceeds a preset reconstruction error threshold, determine that data segment as abnormal to obtain a set of abnormal data segments.

[0028] Specifically, based on the first data stream after dimensionality reduction, a sliding window method is used to segment the continuous time series with a fixed step size, generating a set of data segments with temporal continuity. The variational autoencoder maps the input data segments to a low-dimensional latent space through the encoder network, captures the inherent distribution law of the data by transmitting the states of the hidden layer neurons, and the decoder network reconstructs the original data form according to the latent vectors. The memory unit dynamically records the long-term dependence features of the sequence data during the encoding process, and the hidden state represents the context association information of the current data segment. The reconstruction error is obtained by calculating the root mean square value of the point-by-point difference between the original data segment and the data segment output by the decoder, and the error value reflects the degree to which the data segment deviates from the normal operation mode. The preset reconstruction error threshold is determined based on the error distribution statistics of the historical normal data set. When the error of a certain data segment exceeds the reconstruction error threshold, it indicates that it contains an abnormal feature pattern. The set of abnormal data segments is formed through error threshold screening and spatial clustering analysis, and finally a set of abnormal data segments with clear spatio-temporal markers is output. This step realizes the accurate positioning of weak abnormal signals in a complex operating environment through the non-linear feature extraction ability of the autoencoder and the dynamic threshold determination mechanism.

[0029] In step S13, the vibration signal, the temperature signal, and the current signal are extracted from the set of abnormal data segments, and a convolutional neural network is used to extract local features to obtain a first feature set.

[0030] In a specific implementation manner, the extracting the vibration signal, the temperature signal, and the current signal from the set of abnormal data segments, and using a convolutional neural network to extract local features to obtain a first feature set includes: The vibration signal, the temperature signal, and the current signal are extracted from the set of abnormal data segments, and denoising and normalization processing are performed to obtain a first signal set; According to the first signal set, the signal sequence is divided by a time window method, and a convolutional neural network is used to extract the local features of the signal sequence to obtain a first feature set.

[0031] Specifically, in the abnormal data segment set processing stage, the system synchronously extracts the original waveform data of vibration, temperature, and current signals for the marked abnormal time periods. First, denoising processing based on frequency-domain filtering is performed on the multi-source signals respectively to eliminate the high-frequency interference and power-frequency harmonic components introduced during the sensor acquisition process. Subsequently, the minimum-maximum normalization algorithm is used to unify the dimensions of each signal to the same numerical interval, eliminating the feature weight offset caused by the difference in physical units. The processed signal set is divided into continuous subsequences of fixed length through a sliding time window to ensure the integrity of local time features. The convolutional neural network adopts a multi-layer one-dimensional convolutional kernel structure, and layer-by-layer feature extraction is performed on the signal subsequences through the local perception and weight sharing mechanisms. The first-layer convolutional kernel captures the transient impact waveform of the vibration signal and the gradient change pattern of the temperature signal, and the deep network abstracts the coupled correlation features of current fluctuations and mechanical states. After the feature maps output by the convolutional layer are subjected to non-linear activation and pooling operations, a set of low-dimensional feature vectors representing the local time-frequency characteristics of the signal, that is, the first feature set, is generated. Through multi-level local feature extraction in this step, the original signal is transformed into a structured feature representation with clear physical meanings and suitable for pattern recognition, obtaining the first feature set.

[0032] In step S14, an attention mechanism is used to perform weighted calculation on the first feature set to obtain a weighted feature set.

[0033] In a specific implementation manner, the using an attention mechanism to perform weighted calculation on the first feature set to obtain a weighted feature set includes: Using an attention mechanism to calculate the weight coefficients of the vibration signal, the temperature signal, and the current signal; Performing weighted processing according to the first feature set and the weight coefficients to obtain a weighted feature set.

[0034] Specifically, the system receives the first feature set output from the convolutional neural network, and this set includes vibration feature vectors, temperature feature vectors, and current feature vectors. Among them, the vibration feature vector is generated by extracting the transient impact waveform of the vibration signal through multiple layers of convolution. The temperature feature vector includes the gradient change pattern of the temperature signal, and the current feature vector encodes the coupled correlation characteristics of current fluctuations and mechanical states. The attention mechanism maps the vibration feature vector to a query vector respectively, and the temperature feature vector and the current feature vector to key vectors and value vectors. The cosine similarity calculation is performed between the query vector and the key vector in the feature dimension to obtain the correlation score matrices of the vibration signal and the temperature signal, and the vibration signal and the current signal on the time axis. The correlation score matrices are normalized through Softmax to generate the temperature signal weight coefficient matrix and the current signal weight coefficient matrix, and the dimensions of both are aligned with the time step and feature dimension of the original feature vector.

[0035] The temperature signal weight coefficient matrix and the temperature eigenvector are multiplied element by element to obtain the temperature weighted eigenvector. The current signal weight coefficient matrix and the current eigenvector are multiplied element by element to obtain the current weighted eigenvector. The vibration eigenvector is directly retained as the vibration weighted eigenvector. The finally output weighted feature set is composed of the vibration weighted eigenvector, the temperature weighted eigenvector, and the current weighted eigenvector. Its dimension is exactly the same as that of the input first feature set, but the components in the temperature and current features that are weakly correlated with the time series evolution of the vibration signal are suppressed, and the strongly correlated components are strengthened. This step explicitly captures the driving influence relationship of the vibration signal on the temperature and current features through cross-modal attention modeling, and eliminates the information redundancy between multi-source signals.

[0036] In step S15, according to the weighted feature set, a long short-term memory network is used to extract the second feature set.

[0037] In a specific implementation manner, the using a long short-term memory network to extract the second feature set according to the weighted feature set includes: The long short-term memory network calculates the forgetting gate activation value of the weighted feature set through the forgetting gate; The forgetting gate activation value is compared with a preset activation threshold. If the forgetting gate activation value exceeds the activation threshold, the weighted feature corresponding to the forgetting gate activation value is removed to obtain the second feature set.

[0038] Specifically, the system receives the weighted feature set output from the attention mechanism. This set contains the vibration weighted eigenvector, the temperature weighted eigenvector, and the current weighted eigenvector, where the dimension of each eigenvector is the product of the time step and the feature dimension. The forgetting gate module of the long short-term memory network (LSTM) processes the weighted feature set. The forgetting gate internally contains a linear transformation layer and a Sigmoid activation function. The weighted eigenvectors are input into the forgetting gate step by step in time. The linear transformation layer maps the eigenvector of each time step to the hidden layer space, and the Sigmoid function converts the output of the hidden layer into a forgetting gate activation value between 0 and 1, which represents the redundancy degree of the corresponding feature in the time dimension. The activation threshold preset by the system is determined by the median of the forgetting gate activation values of all eigenvectors in the historical fault data, and the threshold value is stored in the global configuration parameters. When the forgetting gate activation value of a certain time step in the vibration weighted eigenvector, the temperature weighted eigenvector, or the current weighted eigenvector exceeds the threshold, the system sets the eigenvector corresponding to this time step to zero in the second feature set, and the remaining eigenvectors that do not exceed the threshold retain their original values. The finally output second feature set contains the vibration second eigenvector, the temperature second eigenvector, and the current second eigenvector. Its dimension is the same as that of the input weighted feature set, but the redundant features have been dynamically removed. This step quantitatively evaluates the feature time series redundancy through the forgetting gate and realizes the dynamic compression of the feature space.

[0039] In step S16, according to the second feature set, a long short-term memory network is used to analyze the temporal dynamic relationship of the feature sequence, and a third feature set including temporal information is obtained.

[0040] In a specific implementation manner, the step of using a long short-term memory network to analyze the temporal dynamic relationship of the feature sequence according to the second feature set to obtain a third feature set including temporal information includes: Extract the signal sequences of the vibration signal, temperature signal, and current signal from the second feature set; Use the sliding time window method to divide the signal sequences into temporal segments and extract the temporal change features; Calculate the weight values of the temporal segments through the input gate of the long short-term memory network, and perform weighted processing on the temporal change features to obtain weighted temporal features; If the weight value of the temporal segment exceeds the preset weight threshold, filter out the key temporal segments in the temporal segment through the output gate of the long short-term memory network to generate a refined feature set; Use the dynamic time warping algorithm to calculate the morphological similarity of each temporal segment in the refined feature set, and adjust the feature vector dimension of the refined feature set through a gated recurrent unit to obtain a dynamic feature set; According to the dynamic feature set, use the decision tree algorithm to output a third feature set including fault category labels and temporal information.

[0041] Specifically, the system first uses the sliding time window algorithm to segment the multi-source signals on the continuous time axis. Set a window with a fixed length (for example, a 512-sampling-point window with a duration of 2 seconds) and a configurable sliding step (such as a step of 1 / 4 of the window length, that is, sliding once every 0.5 seconds), and cut the continuous signal into partially overlapping temporal segments. The data in each window not only contains the feature values at the current time point but also retains the change trajectories of the signals in the front and rear time windows. For example, the periodic repetition pattern of the gear meshing impact waveform in the vibration signal and the slow-changing rising trend segment of the temperature signal. The parameters of the window length and step are dynamically adjusted according to the physical characteristics of the device. For example, for a motor with a rotational speed of up to 3000 rpm, the window length is shortened to 0.5 seconds to capture more refined transient features.

[0042] The Long Short-Term Memory network (LSTM) receives the time series segments after sliding window segmentation as input. The input gate calculates the associated weights of the features at each time step through the Sigmoid function, and this weight reflects the contribution degree of the current feature to the fault mode prediction. For example, when there is an abnormal fluctuation of a sudden drop in amplitude in a certain time window of the current signal, the input gate will assign a higher weight to this time window. At the same time, the network hidden state is continuously updated to remember long-term dependencies. For example, the cumulative upward trend of the temperature signal in three consecutive windows may trigger an over-temperature warning. If the weighted value of a certain time series segment exceeds a preset threshold (such as 0.7), the output gate marks this segment as a key area and extracts the corresponding hidden state features. This process will screen out the segments containing significant abnormal patterns, such as the phenomenon of periodic enhancement of the waveform caused by bearing damage in the vibration signal.

[0043] Dynamic Time Warping and Morphological Similarity Fusion: For the selected key time series segments, the system uses the Dynamic Time Warping algorithm (DTW) to analyze the morphological co-variation of multi-source signals. For example, the transient drop waveform of the current signal is non-linearly time-aligned with the impact waveform of the vibration signal, and the morphological similarity between the two in terms of morphology is calculated. If there is an offset but morphological similarity (such as both showing a sawtooth waveform with a sharp rise and fall) between the current drop event and the vibration impact on the time axis, it is determined as a fault feature with strong correlation. By constructing an optimal warping path, the time axis distortion error caused by sensor response delay or signal transmission jitter is eliminated. This process can identify synchronous abnormal patterns across sensors. For example, when there is local wear in the gearbox, the co-evolution relationship between the increase in high-frequency vibration energy and the harmonic distortion of the current.

[0044] The Gated Recurrent Unit (GRU) receives the time series features after morphological alignment and dynamically adjusts the feature dimensions through the update gate and the reset gate. The update gate controls the retention ratio of historical feature information. For example, a relatively high update rate is maintained for the long-term trend features of the temperature signal (such as a slow temperature rise lasting for 30 minutes); the reset gate suppresses redundant dimensions. For example, when the energy features of multiple frequency bands in the vibration signal within a certain time window are highly correlated, only the main frequency band features are retained. At this stage, the high-dimensional time series features (such as window features containing 100 dimensions) are compressed into a low-dimensional space (such as 20 dimensions), and at the same time, the discrimination of key dimensions (such as the amplitude of the 3rd harmonic in the current signal) is strengthened through the gating mechanism.

[0045] Finally, the system uses the weighted multi-path decision tree algorithm to integrate the time series features. When constructing classification rules, the decision tree not only divides nodes according to feature values (for example, when the vibration effective value is greater than 5 m / s², enter the left subtree), but also synchronously records the time attributes of the fault features: For each split node, the distribution density of the timing segments that trigger the rule on the time axis is counted. For example, the timestamps corresponding to a certain rule are concentrated in the high-load operation stage of the device (such as from 10 am to 12 pm every day).

[0046] The duration pattern of the fault mode is stored in the leaf node. For example, the characteristic segments of the outer ring fault of the bearing usually show a periodicity of lasting for 2 - 5 seconds and repeating every 30 seconds.

[0047] The decision tree selects the splitting feature according to the principle of maximizing information gain, and preferentially adopts the dimension with clear time correlation (for example, "temperature change rate" is more likely to become a split node than "absolute temperature value"). The third feature set finally output contains two types of information: the structured feature vector (for use by subsequent classifiers) and the associated time metadata (such as the initial trigger timestamp of the fault and the duration of the abnormal mode).

[0048] This process realizes the refined modeling of timing features through multi-layer processing: the sliding window ensures the integrity of local timing features, the LSTM filters key time windows, the dynamic time warping solves the time alignment problem of cross-sensor signals, the GRU optimizes the feature expression efficiency, and the decision tree integrates the time domain statistical characteristics and classification rules.

[0049] In step S17, according to the third feature set, a support vector machine model is used to map the features to predefined fault categories to obtain a preliminary fault classification result.

[0050] Specifically, in the preliminary fault classification stage, the system inputs the third feature set containing timing information into the support vector machine model for pattern mapping. The support vector machine constructs an optimal classification hyperplane to find the decision boundary that can maximize the class interval in the high-dimensional feature space. For the multi-class fault classification requirement, a one-versus-one strategy is adopted to construct multiple binary classifiers for collaborative decision-making. Each classifier focuses on distinguishing the feature distribution differences between a specific fault type and the remaining types. The kernel function non-linearly maps the original features to a more separable hidden space, measures the similarity distance between feature vectors through the radial basis function, and establishes the correspondence between the fault mode and the feature distribution. In the model training stage, the convex optimization problem is solved to determine the support vectors and their weight parameters, forming a classification decision surface with the maximum generalization ability. During the classification process, the similarity between the test sample and each support vector is measured, and the final fault class attribution is determined according to the weighted voting mechanism. This step realizes a robust mapping from the feature space to the fault class through the principle of structural risk minimization, provides a preliminary classification result with clear interpretability for the system, and at the same time retains the classification confidence information for subsequent verification.

[0051] In step S18, the classification confidence is obtained from the preliminary fault classification result. If the confidence is lower than a preset confidence threshold, a long short-term memory network is used to perform secondary classification on the low-confidence samples to obtain the final fault classification result.

[0052] In a specific embodiment, the obtaining the classification confidence from the preliminary fault classification result and, if the confidence is lower than the preset confidence threshold, using a long short-term memory network to perform secondary classification on the low-confidence samples to obtain the final fault classification result includes: According to the preliminary fault classification result, the softmax function is used to calculate the classification confidence of each fault category; If the highest confidence in the classification confidence is lower than the confidence threshold determined in advance through cross-validation, the corresponding third feature is marked as a low-confidence sample; According to the low-confidence samples, the vibration signal and the temperature signal are extracted from the time-series information of the third feature set, and a long short-term memory network is used to perform feature sequence modeling on the vibration signal and the temperature signal to obtain an enhanced feature sequence; Through the enhanced feature sequence, a deep feature fusion classifier is used to perform secondary classification on the low-confidence samples and output a fault classification result after probability calibration.

[0053] Specifically, in the classification result verification stage, the system performs confidence evaluation and calibration on the preliminary fault classification result output by the support vector machine. Based on the geometric distance characteristics of the classification decision surface, the decision function values of each fault category are converted into a normalized probability distribution through the Softmax function, and the maximum value corresponds to the confidence level of the sample belonging to this category. When the highest confidence is lower than the preset threshold, the system determines that the sample is in the classification boundary fuzzy area and triggers the secondary classification mechanism. For the marked low-confidence samples, the time-series feature vectors of the vibration and temperature signals are separated from the third feature set, and a multi-level time-series dependence model is constructed through a stacked long short-term memory network. The deep-layer units of the network capture the long-period patterns and the correlation relationships of mutation events in the signal fluctuations and output an enhanced feature sequence that fuses multiple time scales. The deep feature fusion classifier adopts a dual-channel input architecture, cross-domain splices the static features of the current signal of the original third feature set and the enhanced time-series features, performs non-linear feature interaction modeling through a fully connected layer, and finally adjusts the output distribution through a probability calibration layer to eliminate the probability bias of the initial classification. This step effectively solves the problem of misjudgment of boundary samples under complex working conditions and improves the system's ability to identify fuzzy fault patterns by establishing a hierarchical verification mechanism and a spatio-temporal feature complementary model.

[0054] Refer to Figure 2 , the second embodiment of the present invention provides a spinning machine fault detection system based on deep learning, including: A data acquisition module, configured to acquire a high-dimensional heterogeneous data stream from a spinning machine, including vibration signals, temperature signals, and current signals, and retain the main eigenvectors by using principal component analysis to obtain a first data stream after dimensionality reduction; An abnormal data analysis module, configured to calculate a reconstruction error based on the first data stream through a memory cell state and a hidden state, and if the reconstruction error exceeds a preset reconstruction error threshold, determine that the data segment of the reconstruction error is abnormal to obtain a set of abnormal data segments; A first feature analysis module, configured to extract the vibration signal, the temperature signal, and the current signal from the set of abnormal data segments, and extract local features by using a convolutional neural network to obtain a first feature set; A weighted feature analysis module, configured to perform weighted calculation on the first feature set by using an attention mechanism to obtain a weighted feature set; A second feature analysis module, configured to extract a second feature set by using a long short-term memory network based on the weighted feature set; A third feature analysis module, configured to analyze the temporal dynamic relationship of a feature sequence by using a long short-term memory network based on the second feature set to obtain a third feature set containing temporal information; A preliminary fault classification module, configured to map features to predefined fault categories by using a support vector machine model based on the third feature set to obtain a preliminary fault classification result; A final fault classification module, configured to obtain a classification confidence level from the preliminary fault classification result, and if the confidence level is lower than a preset confidence level threshold, perform secondary classification on low-confidence samples by using a long short-term memory network to obtain a final fault classification result.

[0055] It should be noted that a fault detection device for a spinning machine based on deep learning provided in an embodiment of the present invention is used to execute all process steps of a fault detection method for a spinning machine based on deep learning in the above embodiment, and the working principles and beneficial effects of the two correspond one by one, so details will not be repeated here.

[0056] An embodiment of the present invention further provides an electronic device. The electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a fault detection program for a spinning machine based on deep learning. When the processor executes the computer program, the steps in the above-mentioned embodiments of various fault detection methods for a spinning machine based on deep learning are implemented, such as Figure 1 Step S11 shown. Alternatively, when the processor executes the computer program, the functions of each module / unit in the above-mentioned device embodiments are implemented, such as a fault detection module for a spinning machine based on deep learning.

[0057] Exemplarily, the computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program in the electronic device.

[0058] The electronic device may be a computing device such as a desktop computer, a notebook, a palm computer, and a smart tablet. The electronic device may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the above components are only examples of the electronic device and do not constitute a limitation on the electronic device. It may include more or fewer components than the above, or combine certain components, or different components. For example, the electronic device may further include input / output devices, network access devices, a bus, etc.

[0059] The so-called processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the electronic device and connects various parts of the entire electronic device through various interfaces and circuits.

[0060] The memory can be used to store the computer programs and / or modules. By running or executing the computer programs and / or modules stored in the memory and invoking the data stored in the memory, the processor can implement various functions of the electronic device. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory can include high-speed random access memory and can also include non-volatile memory, such as a hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one magnetic disk storage device, flash device, or other volatile solid-state storage devices.

[0061] Among them, if the modules / units integrated in the electronic device are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-mentioned embodiment methods of the present invention, it can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0062] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided by the present invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0063] The specific embodiments described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. In particular, for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A fault detection method for a spinning machine based on deep learning, characterized in that, Including: Obtain a high-dimensional heterogeneous data stream from a spinning machine, including vibration signals, temperature signals, and current signals. Use principal component analysis to retain the main eigenvectors and obtain a first data stream after dimensionality reduction; According to the first data stream, calculate the reconstruction error through the memory cell state and the hidden state. If the reconstruction error exceeds a preset reconstruction error threshold, determine the data segment with the reconstruction error as abnormal and obtain a set of abnormal data segments; Extract the vibration signal, the temperature signal, and the current signal from the set of abnormal data segments, and use a convolutional neural network to extract local features to obtain a first feature set; Use an attention mechanism to perform weighted calculation on the first feature set to obtain a weighted feature set; According to the weighted feature set, use a long short-term memory network to extract a second feature set; According to the second feature set, use a long short-term memory network to analyze the temporal dynamic relationship of the feature sequence to obtain a third feature set containing temporal information; According to the third feature set, use a support vector machine model to map the features to predefined fault categories to obtain a preliminary fault classification result; Obtain the classification confidence from the preliminary fault classification result. If the confidence is lower than a preset confidence threshold, use a long short-term memory network to perform secondary classification on the low-confidence samples to obtain the final fault classification result.

2. The method for detecting faults of a spinning machine based on deep learning according to claim 1, wherein, The calculating the reconstruction error through the memory cell state and the hidden state according to the first data stream, and if the reconstruction error exceeds a preset reconstruction error threshold, determining the data segment with the reconstruction error as abnormal to obtain a set of abnormal data segments includes: Obtain continuous data segments from the first data stream, and divide the continuous data segments by a sliding window method to obtain divided data segments; Use a variational autoencoder to encode and decode the divided data segments, and calculate the reconstruction error through the memory cell and the hidden state to obtain a set of reconstruction errors; If the reconstruction error of any data segment in the set of reconstruction errors exceeds a preset reconstruction error threshold, determine that data segment as abnormal to obtain a set of abnormal data segments.

3. The method for detecting faults of a spinning machine based on deep learning according to claim 1, wherein, The extracting the vibration signal, the temperature signal, and the current signal from the set of abnormal data segments, and using a convolutional neural network to extract local features to obtain a first feature set includes: Extract the vibration signal, the temperature signal, and the current signal from the set of abnormal data segments, and perform denoising and normalization processing to obtain a first signal set; According to the first signal set, divide the signal sequence by a time window method, and use a convolutional neural network to extract the local features of the signal sequence to obtain a first feature set.

4. The method for detecting faults of a spinning machine based on deep learning according to claim 1, wherein, The using an attention mechanism to perform weighted calculation on the first feature set to obtain a weighted feature set includes: Use an attention mechanism to calculate the weight coefficients of the vibration signal, the temperature signal, and the current signal; Perform weighted processing according to the first feature set and the weight coefficients to obtain a weighted feature set.

5. The method for detecting faults of a spinning machine based on deep learning according to claim 1, wherein The extracting a second feature set using a long short-term memory network according to the weighted feature set includes: The long short-term memory network calculates the forget gate activation value of the weighted feature set through the forget gate; Compare the forget gate activation value with a preset activation threshold. If the forget gate activation value exceeds the activation threshold, eliminate the weighted features of the forget gate activation value to obtain a second feature set.

6. The method for detecting faults of a spinning machine based on deep learning according to claim 1, wherein Based on the second feature set, analyze the temporal dynamic relationship of the feature sequence using a long short-term memory network to obtain a third feature set containing temporal information, including: Extract the signal sequences of the vibration signal, temperature signal, and current signal from the second feature set; Use the sliding time window method to divide the signal sequence into temporal segments and extract temporal change features; Calculate the weight values of the temporal segments through the input gate of the long short-term memory network and perform weighted processing on the temporal change features to obtain weighted temporal features; If the weight value of the temporal segment exceeds a preset weight threshold, filter out the key temporal segments in the temporal segment through the output gate of the long short-term memory network to generate a refined feature set; Use the dynamic time warping algorithm to calculate the morphological similarity of each temporal segment in the refined feature set, and adjust the feature vector dimension of the refined feature set through a gated recurrent unit to obtain a dynamic feature set; Based on the dynamic feature set, use the decision tree algorithm to output a third feature set containing fault category labels and temporal information.

7. The method for detecting faults of a spinning machine based on deep learning according to claim 1, characterized in that Obtain the classification confidence from the preliminary fault classification result. If the confidence is lower than a preset confidence threshold, use the long short-term memory network to perform secondary classification on the low-confidence samples to obtain the final fault classification result, including: Based on the preliminary fault classification result, use the softmax function to calculate the classification confidence of each fault category; If the highest confidence in the classification confidence is lower than the confidence threshold determined in advance through cross-validation, mark the corresponding third feature as a low-confidence sample; Based on the low-confidence samples, extract the vibration signal and temperature signal from the temporal information of the third feature set, and use the long short-term memory network to perform feature sequence modeling on the vibration signal and the temperature signal to obtain an enhanced feature sequence; Through the enhanced feature sequence, use a deep feature fusion classifier to perform secondary classification on the low-confidence samples and output a fault classification result with probability calibration.

8. A fault detection system for a spinning machine based on deep learning, characterized in that, Including: A data acquisition module for obtaining a high-dimensional heterogeneous data stream from a spinning machine, including a vibration signal, a temperature signal, and a current signal, and retaining the main feature vectors using principal component analysis to obtain a first data stream after dimensionality reduction; An abnormal data analysis module for calculating the reconstruction error based on the first data stream through the memory cell state and the hidden state. If the reconstruction error exceeds a preset reconstruction error threshold, determine that the data segment of the reconstruction error is abnormal to obtain a set of abnormal data segments; A first feature analysis module for extracting the vibration signal, the temperature signal, and the current signal from the set of abnormal data segments and using a convolutional neural network to extract local features to obtain a first feature set; A weighted feature analysis module for performing weighted calculation on the first feature set using an attention mechanism to obtain a weighted feature set; A second feature analysis module, configured to extract a second feature set by using a long short-term memory network according to the weighted feature set; A third feature analysis module, configured to analyze the temporal dynamic relationship of the feature sequence by using a long short-term memory network according to the second feature set, so as to obtain a third feature set including temporal information; A preliminary fault classification module, configured to map features to predefined fault categories by using a support vector machine model according to the third feature set, so as to obtain a preliminary fault classification result; A final fault classification module, configured to obtain a classification confidence level from the preliminary fault classification result. If the confidence level is lower than a preset confidence level threshold, a long short-term memory network is used to perform secondary classification on samples with low confidence levels, so as to obtain a final fault classification result.

9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the method for detecting faults of a spinning machine based on deep learning according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein when the computer program runs, the device where the computer-readable storage medium is located is controlled to execute the method for detecting faults of a spinning machine based on deep learning according to any one of claims 1 to 7.

Images