Industrial bearing vibration time sequence signal fault prediction method and system fusing attention mechanism and LSTM

Abstract

The invention discloses an attention mechanism and LSTM fused industrial bearing vibration time sequence signal fault prediction method and system. The method comprises the following steps: collecting a bearing vibration signal and carrying out filtering, noise reduction and normalization preprocessing; constructing a deep learning model combining the bidirectional BiLSTM and a coordinate attention mechanism to extract bidirectional time sequence features and enhance key fault features; carrying out model training by adopting a multi-target composite loss function and an Adam optimizer, and introducing an early stop mechanism to prevent overfitting; performing fault type identification and degree evaluation on the real-time vibration signal by using the trained model, and performing quantitative analysis by fusing multi-scale spectrum kurtosis features and nonlinear kinetic parameters; and finally, outputting a fault diagnosis report, and triggering multi-stage early warning based on an adaptive threshold. The method can realize high-precision and high-reliability bearing fault prediction and health state evaluation, and is suitable for intelligent operation and maintenance of industrial equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment state monitoring and fault prediction, and particularly relates to an intelligent fault prediction method and system for industrial bearing vibration signals based on deep learning. BACKGROUND

[0002] Industrial bearings, as the core components of rotating machinery, their running state is directly related to the safety and reliability of the entire equipment and even the production line. Therefore, real-time and accurate fault prediction and health management (PHM) of bearings have great economic and safety value. Vibration signal analysis is the most commonly used and effective means in bearing state monitoring, because the occurrence and development of faults will directly change the vibration characteristics and produce specific impact responses.

[0003] Traditional bearing fault diagnosis methods are heavily dependent on signal processing techniques and manual feature extraction, such as extracting time domain and frequency domain features of signals through fast Fourier transform (FFT), wavelet transform, envelope spectrum analysis, etc., and then combining expert experience or machine learning classifiers (such as support vector machine SVM) for fault recognition. However, these methods have obvious limitations: first, manual feature extraction is heavily dependent on domain expertise, and the effectiveness and universality of features are difficult to guarantee; second, traditional methods are difficult to capture complex nonlinear, non-stationary characteristics and long-range temporal dependencies in vibration signals, and lack sensitivity to early weak faults and complex faults.

[0004] In recent years, deep learning technology, especially long short-term memory network (LSTM), has shown great potential in the field of fault prediction due to its strong time series modeling capability. LSTM can automatically learn the temporal patterns and dependencies in vibration signals. However, the standard LSTM model may dilute key information or have difficulty focusing on fault-sensitive periods when processing long sequences. The introduction of attention mechanism can effectively alleviate this problem, as it assigns different weights to different parts of the sequence, allowing the model to focus on the most relevant features to the fault. However, existing methods mostly use conventional attention mechanisms, which still have room for improvement in their ability to collaboratively focus on the spatial and temporal features of vibration signals.

[0005] In addition, existing intelligent diagnosis models mostly focus only on fault type classification, ignoring the equally important fault severity quantification evaluation, making it difficult to meet the precise needs of predictive maintenance for fault evolution trend judgment. At the same time, the training process of the model faces challenges such as balancing classification and regression multi-task learning, overfitting, and consistency between prediction results and physical mechanisms.

[0006] In summary, developing an intelligent diagnosis method capable of automatically extracting deep temporal features, accurately focusing on key fault information, and simultaneously realizing fault type identification and severity quantification evaluation has become a key problem that the industry urgently needs to solve. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings of the prior art, provide a bearing fault prediction method and system capable of automatically extracting key features, simultaneously realizing fault type identification and severity quantification evaluation, and having high reliability.

[0008] In a first aspect, the embodiments of the present application provide an industrial bearing vibration time series signal fault prediction method fusing attention mechanism and LSTM, which comprises: S1, collecting vibration time domain signals in the running process of an industrial bearing by using a vibration sensor, and sequentially performing filtering and noise reduction and normalization processing on the collected signals to obtain standardized vibration time series data; S2, constructing a deep learning model comprising a bidirectional LSTM layer and a coordinate attention mechanism layer, wherein the bidirectional LSTM layer is used to extract bidirectional time series features in the vibration time series data, and the coordinate attention mechanism layer is used to perform weighting processing on the hidden state output by the bidirectional LSTM layer in time and feature dimension, thereby enhancing the attention degree to the key fault features; S3, training the deep learning model using a labeled bearing vibration data set, optimizing the parameters using an Adam optimizer, using a cross-entropy error function as the loss function, and introducing an early stopping mechanism to prevent overfitting; S4, inputting the real-time collected vibration data into the trained deep learning model, and simultaneously outputting a fault type identification result and a fault degree evaluation value, wherein the fault degree evaluation is quantitatively calculated based on the fault impact features learned by the model; S5, generating a diagnosis report containing both the fault type and the fault degree, and triggering a multi-level warning mechanism when the fault degree evaluation value exceeds an adaptive threshold value set based on historical data.

[0009] In a second aspect, the embodiments of the present application provide an industrial bearing vibration time series signal fault prediction system fusing attention mechanism and LSTM, which is applied to the industrial bearing vibration time series signal fault prediction method fusing attention mechanism and LSTM as described in the first aspect, and comprises: A signal acquisition and preprocessing module is configured to collect vibration time domain signals in the running process of an industrial bearing by using a vibration sensor, and to preprocess the collected signals, including filtering and noise reduction and normalization processing, to obtain standardized vibration time series data. A model construction module is configured to collect vibration time domain signals of an industrial bearing during operation by using a vibration sensor, and sequentially perform filtering and noise reduction and normalization processing on the collected signals to obtain standardized vibration time series data. A model training module is configured to train the deep learning model using a labeled bearing vibration data set, perform parameter optimization by using an Adam optimizer, use a cross-entropy error function as a loss function, and introduce an early stopping mechanism to prevent overfitting. A real-time fault diagnosis module is configured to input real-time collected vibration data into the trained deep learning model, and synchronously output a fault type recognition result and a fault degree evaluation value, wherein the fault degree evaluation is quantitatively calculated based on fault impact features learned by the model. A diagnosis report and early warning module is configured to generate a diagnosis report containing a fault type and a fault degree, and trigger a multi-level early warning mechanism when the fault degree evaluation value exceeds an adaptive threshold value set based on historical data.

[0010] In a third aspect, an electronic device is provided, including: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the industrial bearing vibration time series signal fault prediction method of the fusion attention mechanism and LSTM when executing the instructions.

[0011] In a fourth aspect, a computer-readable storage medium is provided, which stores a program instructing a device to execute the industrial bearing vibration time series signal fault prediction method of the fusion attention mechanism and LSTM.

[0012] The industrial bearing vibration time series signal fault prediction method and system of the fusion attention mechanism and LSTM have the following remarkable beneficial effects: 1. The extraction and focusing ability of fault features is improved: by combining bidirectional BiLSTM and coordinate attention mechanism (CA), the model can not only capture the bidirectional long-range time sequence dependency relationship in the vibration signal, but also adaptively enhance the attention to fault-sensitive key information in the time and feature channel dimensions, effectively suppress irrelevant noise interference, and significantly improve the extraction ability of early weak fault features.

[0013] 2. The synchronous and accurate diagnosis of fault type and severity is realized: the model adopts a multi-task learning architecture, can output high-precision fault type classification results and continuous fault degree regression estimation values in parallel, solves the problem that traditional methods are limited to fault classification and cannot quantitatively evaluate the severity, and provides a more comprehensive decision basis for predictive maintenance.

[0014] 3. Enhanced generalization and reliability of the model: By training with a dynamically weighted multi-objective loss function, an adaptive optimization strategy, and a KL divergence constraint combined with physical mechanisms, the model training process is more stable, and the prediction results are not only data-driven but also conform to the physical characteristics of fault impact, significantly improving the generalization ability of the model and the reliability of the diagnosis results.

[0015] 4. Intelligent and automated diagnosis and early warning closed loop: The system can automatically generate a structured diagnosis report and set adaptive thresholds for multi-level early warning based on historical data, while providing specific maintenance decision recommendations, realizing a complete and automated process from state perception, intelligent diagnosis to decision support, greatly improving the efficiency and intelligent level of equipment health management. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The fusion attention mechanism and LSTM-based industrial bearing vibration time series signal fault prediction method flowchart provided by an embodiment of the present application is shown in the figure.

[0017] Figure 2 The fusion attention mechanism and LSTM-based industrial bearing vibration time series signal fault prediction system architecture diagram provided by the present application is shown in the figure.

[0018] Figure 3 The electronic device schematic diagram provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all.

[0020] It should be noted that at least one in the embodiments of the present application refers to one or more, and multiple refers to two or more. Unless otherwise defined, all technical and scientific terms used in the present application have the same meaning as understood by those skilled in the art to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing the specific embodiments of the present application, and are not intended to limit the present application.

[0021] Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0022] Embodiment one

[0023] Figure 1 The fusion attention mechanism and LSTM-based industrial bearing vibration time series signal fault prediction method flowchart provided by an embodiment of the present application is shown in the figure.Figure 1 As shown in the figure, an industrial bearing vibration time series signal fault prediction method combining attention mechanism and LSTM includes: S1, collect vibration time domain signals during the operation of the industrial bearing using a vibration sensor, filter and denoise the collected signals in turn, and normalize the signals to obtain standardized vibration time series data. The standardized processing of the original vibration signal is realized. The vibration time domain signal of the industrial bearing in operation is collected by the sensor, and is filtered and denoised (the fault feature frequency band is retained and noise is suppressed) and normalized (the data scale is unified) in turn to generate high-quality standardized time series data, providing effective input for subsequent models.

[0024] Specifically, in this embodiment, the preprocessing in step S1 includes the following sub-steps: S1.1, signal filtering: an adaptive band-pass filter is used, and the passband frequency range is 0.5 times to 3 times the bearing characteristic frequency, so as to retain the fault characteristic frequency component and suppress irrelevant frequency band interference; irrelevant frequency interference is filtered out, and the key frequency component related to the bearing fault is retained. If the bearing characteristic frequency is 100 Hz, the filter only allows signals between 50 Hz and 300 Hz to pass through, effectively highlighting the fault impact signal.

[0025] S1.2, noise suppression processing: a threshold denoising algorithm based on wavelet packet is used, db4 wavelet basis is used for multi-layer wavelet packet decomposition, and a soft threshold value is used for processing the detail coefficients; reduce random noise in the signal and improve the signal-to-noise ratio. Using wavelet packet decomposition technology, small noise fluctuations in the signal are removed like removing coffee stains, making the fault characteristics clearer.

[0026] S1.3, signal normalization processing: the maximum-minimum normalization method is used to normalize the signal amplitude to a preset numerical range; the data scale is unified to speed up the model training convergence. All vibration signal amplitudes are compressed to the range of [0, 1] or [-1, 1], avoiding the influence of certain values that are too large or too small on model learning.

[0027] S1.4, data enhancement processing: expand the training data set through time series data enhancement technology, which includes one or more of time warping, amplitude scaling, and adding Gaussian noise; expand the training data volume and improve the model generalization ability. For example, stretch or compress the original signal or add slight noise to generate more samples and make the model learn more fully.

[0028] S1.5, feature extraction pre-processing: calculate the time domain feature parameters of the vibration signal, which include one or more of the root mean square value, kurtosis, peak factor and pulse index, and use these features and pre-processed time series data as inputs to the fault feature extraction model. Extracting traditional time domain features and fusing them with the original signal to input the model. For example, calculating the roughness (kurtosis) and impact strength (pulse index) of the signal provides more rich fault information for the model.

[0029] These five sub-steps are progressive, together ensuring that the data input to the model is clean, standardized and rich, laying a solid foundation for subsequent accurate diagnosis. Just like preparing ingredients for a chef: first, filter, then wash (denoise), cut and arrange (normalize), increase ingredients (data augmentation), and finally, simple pre-processing (feature extraction), so that you can fry a good dish (accurate diagnosis).

[0030] S2, build a deep learning model containing a bidirectional LSTM layer and a coordinate attention mechanism layer, wherein the bidirectional LSTM layer is used to extract bidirectional time series features in the vibration time series data, and the coordinate attention mechanism layer is used to weight process the hidden state output by the bidirectional LSTM layer in time and feature dimensions, to enhance the attention to key features. Build a deep learning model with feature extraction and key information enhancement capabilities. Use a bidirectional BiLSTM layer to capture the bidirectional long-term time series dependency in the vibration signal, and use a coordinate attention mechanism layer to weight process the BiLSTM output in time and feature dimensions, to strengthen the expression of fault sensitive features and suppress irrelevant information interference.

[0031] Specifically, in the present embodiment, the deep learning model constructed in step S2 is a hierarchical feature extraction and fusion network, specifically including: S2.1, bidirectional time series feature extraction layer: use a multi-layer bidirectional long short-term memory network BiLSTM as the core time series feature extractor, and the hidden state of each time step is used to capture the long-range bidirectional dependency relationship related to the fault in the vibration signal; the bidirectional time series feature extraction layer extracts past and future context information from the vibration signal like reading an article forward and backward, capturing the long-term dependency relationship of the fault. For example, the BiLSTM network analyzes the front and back paragraphs of the vibration signal at the same time, and identifies abnormal periodic impact patterns (such as a collision sound that occurs once every revolution).

[0032] The bidirectional time series feature extraction layer in step S2.1 satisfies the following conditions: The number of layers of the multi-layer bidirectional long short-term memory network BiLSTM is 2 to 4 layers, and the number of hidden units of each layer is configurable between 64 and 256. The network depth and capacity range are defined by the structure configuration to balance the model expression and computational efficiency. Specifically, the number of BiLSTM layers: 2 to 4 layers, to avoid insufficient feature extraction or overfitting. The number of hidden units: 64 to 256, to provide sufficient feature expression capability.

[0033] The BiLSTM layer receives the standardized vibration time series data preprocessed by step S1, and the forward and backward LSTM units process the input sequence in time order and reverse order respectively, and splice the hidden states of each time step in two directions to form a complete hidden state output that integrates the bidirectional context information. The bidirectional feature fusion mechanism completely captures the time context information and enhances the fault feature expression capability. The forward LSTM processes the sequence in time order to capture historical dependencies; the backward LSTM processes the sequence in reverse order to capture future dependencies; and the spliced bidirectional hidden states form a complete feature representation that integrates bidirectional information.

[0034] The BiLSTM network adopts weight normalization combined with activation value scaling technology during training, which reparameterizes the convolution weight matrix to meet the distribution characteristics of zero mean and unit variance, and introduces a learnable scaling parameter before the activation function to stabilize the activation distribution during network training, accelerate convergence and alleviate the internal covariate shift problem. Training stability optimization technology solves the internal covariate shift problem in deep network training and accelerates convergence. It includes weight normalization: reparameterize the weight matrix to make it have a distribution of zero mean and unit variance; activation value scaling: introduce a learnable scaling parameter before the activation function to stabilize the activation distribution.

[0035] The complete hidden state output is not only passed to the subsequent coordinate attention mechanism layer, but also the output of the last time step is additionally used as a global time context feature, which is spliced with the feature vector after attention weighted summation and input to the classification and regression output layer to enhance the model's perception of the overall state of the sequence. The global context enhancement strategy strengthens the model's perception of the overall state of the sequence and avoids excessive localization of the attention mechanism. The output of the last time step of BiLSTM, the global time context feature, and the feature vector after attention weighted summation are spliced and input to the output layer, while utilizing local key information and global state information.

[0036] Assume that a bearing vibration signal is analyzed: bidirectional LSTM: simultaneously observe the vibration pattern 0.5 seconds before the current time and 0.5 seconds after the current time, and judge whether a fault impact occurs; weight normalization: like a standard food processing standard, ensure that the processing results of each layer of the network are stable and distributed, and avoid subsequent layer learning difficulties; global context concatenation: when judging the fault, not only pay attention to the most abnormal waveform segment (attention focus), but also refer to the overall state of the entire signal segment (the last time step output), to avoid misjudgment. Through specific structure design and technical means, ensure that the bidirectional time sequence feature extraction layer achieves optimal performance and stability in the industrial vibration signal scene.

[0037] S2.2, coordinate attention mechanism layer: receives the hidden state sequence of all time steps of the BiLSTM layer, and performs attention weight calculation along the time dimension and the feature dimension through the coordinate attention mechanism, to generate a two-dimensional attention weight matrix, and recalibrate and weight the BiLSTM output features, to strengthen the expression of fault-sensitive time periods and feature channels; the coordinate attention mechanism layer is like a searchlight, which automatically focuses on the most critical fault signal segment and feature channel in the time and feature dimensions. The model will pay special attention to the time point with the highest impact energy (when) and the feature channel that best represents the inner ring fault (where), and enhance their weights.

[0038] Specifically, the implementation process of the coordinate attention mechanism layer in step S2.2 is as follows: Receive the complete hidden state sequence output by the BiLSTM layer, first perform one-dimensional global pooling operation along the time dimension and the feature dimension, generate global context descriptors in the time dimension and the feature dimension, perform global information compression on the feature sequence output by the BiLSTM, capture the global statistical characteristics of the entire sequence in the time dimension and the feature dimension, and provide a basis for subsequent weight calculation. Specifically, along the time dimension global pooling: average all time steps of each feature channel to generate a feature dimension descriptor , representing the global importance of each feature channel. Along the feature dimension global pooling: average all feature channels of each time step to generate a time dimension descriptor , representing the global importance of each time step.

[0039] Concatenate and fuse the descriptors in the two dimensions, then generate intermediate feature maps through a convolution layer with shared weights and a nonlinear activation function, promote the interaction of time information and channel information, and generate more rich intermediate features through nonlinear transformation.

[0040] Separate the intermediate feature maps along the original dimensions to obtain a time attention weight vector and a channel attention weight vector, respectively, and decouple the attention weights that act on the time dimension and the feature dimension from the fused features.

[0041] The time attention weight vector and the channel attention weight vector are operated to generate a two-dimensional attention weight matrix. The two-dimensional attention weight matrix is multiplied element by element with the complete hidden state sequence output by the BiLSTM, which enhances the fault-sensitive time period and key feature channel, and suppresses irrelevant information. The one-dimensional time weight and the channel weight are combined into a two-dimensional attention map, which accurately calibrates the original features.

[0042] Specifically, the coordinate attention mechanism layer in step S2.2 specifically includes the following processing process: given the complete hidden state sequence output by the BiLSTM layer , where T is the total number of time steps, D is the feature dimension, and one-dimensional global average pooling is performed along the time dimension and the feature dimension respectively to generate a time dimension global context descriptor and a feature dimension global context descriptor : , , wherein is the jth column of the feature matrix H. It represents the value of the jth feature channel at all time steps (:). is summed along the feature dimension D. All D elements in the jth column are summed. is the time dimension global context descriptor. A length T vector, where each element represents the average response intensity of the ith time step on all feature channels. It encodes the global importance of each time step. is the ith row of the feature matrix H. It represents the value of all D feature channels at the ith time step. It is a length D vector. is summed along the time dimension T. All T elements in the ith row are summed. is the feature dimension global context descriptor. A length D vector, where each element represents the average response intensity of the jth feature channel at all time steps. It encodes the global importance of each feature channel.

[0043] The two-dimensional spatio-temporal features H are compressed into two one-dimensional global descriptor vectors respectively, one capturing the global information of the time dimension , and the other capturing the global information of the feature channel dimension , to prepare for subsequent calculation of attention weights.

[0044] After splicing the two descriptors, a 1x1 convolution with shared weights, batch normalization and a nonlinear activation function are used to generate an intermediate feature map : , where, is a Sigmoid function, a nonlinear activation function, compressing the output value between 0 and 1, introducing nonlinearity, is a batch normalization operation, is a one-dimensional convolution with kernel size 1, using a one-dimensional convolution layer with kernel size 1 to process the concatenated vector. The role of 1x1 convolution is to interact and fuse across features, and can reduce or increase dimension; is a concatenation operation. The time descriptor (length T) and the channel descriptor (length D) are concatenated in the same dimension to form a longer vector with length T+D. is a batch normalization operation. The output after convolution is standardized to speed up training and improve stability. F is the intermediate feature map. A feature vector after transformation and activation, whose dimension is determined by the output channel number C of the 1x1 convolution, with a shape of [C, T+D]. It fuses the global information of time and channel.

[0045] Split F along the original dimension into time attention weights and channel attention weights , and normalize them through the Sigmoid function respectively: , ; where, is the time dimension part of the intermediate feature map F. The first T elements of F are split out. is the feature dimension part of the intermediate feature map F. The last D elements of F are split out. is the time attention weight vector. A vector of length T, each element is a weight between 0 and 1, indicating the importance of the i-th time step to fault diagnosis. The closer the value is to 1, the more important it is. is the channel attention weight vector. A vector of length D, each element is a weight between 0 and 1, indicating the importance of the j-th feature channel to fault diagnosis. The closer the value is to 1, the more important it is.

[0046] Decouple the time and channel information from the fused feature F, and generate attention weights that act on the two dimensions respectively.

[0047] Generate a two-dimensional attention weight matrix : , is the outer product operation. The vector (Tx1) and vector (1xD) multiplication, generating a matrix M(TxD). is a two-dimensional attention weight matrix. Each element in the matrix represents the comprehensive importance weight of the i-th time step and the j-th feature channel.

[0048] Element-wise multiplication of the weight matrix M and the original hidden state H gives the calibrated and weighted feature representation : , where ⊙ represents element-wise multiplication (Hadamard product) operation, which multiplies the original feature matrix H(TxD) and the attention weight matrix M(TxD) at the corresponding positions. represents the weighted and calibrated feature sequence. Its shape is the same as H, both [T,D]. After attention weighting, the values of important time steps and important feature channels are enhanced, while the values of unimportant time steps and feature channels are suppressed.

[0049] Finally, the weighted feature is used for subsequent feature fusion and fault diagnosis output. Finally, a feature representation focusing on key information of faults is obtained, which is input into subsequent network layers for fault classification and degree regression, thereby improving the diagnostic performance and interpretability of the model.

[0050] Suppose the BiLSTM output is a sequence of 100 time steps (T=100) and 64-dimensional features (D=64). Step 1: generate a 100-dimensional vector and a 64-dimensional vector respectively. Steps 2&3: transform and split the two vectors into a new 100-dimensional weight (accurately corrected time importance) and a new 64-dimensional weight (accurately corrected channel importance). For example, may show that the 25th time step has the highest weight, indicating that this time is most likely the fault impact point. Step 4: combine and into a 100x64 matrix M, and then use this matrix to amplify the values at the 25th time step and those important feature channels determined, finally obtaining the enhanced feature .

[0051] The core value of this scheme is to define an efficient and accurate two-dimensional attention mechanism calculation process, which is a key technology to improve the fault diagnosis performance and interpretability of the model.

[0052] S2.3, Feature fusion and compression layer: the coordinate attention weighted feature sequence is weighted and summed in the time dimension, compressed into a fixed-dimensional feature vector, which fuses the key information most relevant to the fault in the entire sequence; the feature fusion and compression layer condenses and refines the rich information of the entire time sequence into a fixed-length feature vector containing the essence. For example, the analysis result of a 10-second vibration signal is compressed into a core feature vector containing the fault type as inner ring crack and the severity as 0.7.

[0053] The feature fusion and compression layer in the step S2.3 is implemented as follows: Receiving the coordinate attention weighted feature sequence where T is the total number of time steps, and D is the feature dimension; receiving the calibrated important features. As the input of this step, the feature sequence that has been weighted by the coordinate attention mechanism is where the unimportant time and feature channel information has been suppressed, and the important information has been enhanced.

[0054] Using an attention weight based time sequence fusion method, the weighted feature sequence is compressed in the time dimension to generate a fixed-dimensional global feature vector : , where, represents the weighted feature vector of the t-th time step, is the fusion weight coefficient of the t-th time step; the information of the entire time sequence is adaptively (rather than averagely) compressed into a vector. This method can give different attention degrees according to the importance of each time step to the current fault diagnosis task. represents the fusion weight coefficient of the t-th time step. This coefficient is not preset, but is calculated by a learnable scoring function, ensuring that the model only focuses on the time steps containing key fault information. The fusion weight coefficient is calculated by the time step importance scoring function

[0055] and normalized by the Softmax function: , , , where, is the weight matrix, and are learnable parameters, ​is the hyperbolic tangent activation function; the importance score is calculated at each time step and converted to a weight. This is a learnable attention process that enables the model to learn which parts of the sequence are most crucial for the final decision.

[0056] The resulting global feature vector fuses the most relevant information from the entire time series that is most relevant to the fault and serves as input to the subsequent classification and regression output layers. The feature representation needed to generate the final decision. The output is a fixed-dimensional global feature vector that is no longer time-series data but a condensed feature sample that encapsulates the essence of the entire sequence and can be directly input to fully connected layers for classification and regression.

[0057] S2.4, Classification and Regression Output Layers: Receive the fixed-dimensional feature vector and output the classification probabilities of fault types through a fully connected layer and a Softmax function, while outputting the regression estimate of fault severity through another fully connected layer and a linear activation function. The classification and regression output layers, like expert consultations, simultaneously provide qualitative judgments (what fault) and quantitative assessments (how severe). For example, one branch outputs fault type probabilities (e.g., inner race crack: 85%, ball shedding: 10%, normal: 5%), and the other branch directly outputs severity scores (e.g., 0.72 / 1.0).

[0058] The classification and regression output layers in step S2.4 are implemented as follows: Receive the fixed-dimensional global feature vector , which is simultaneously input to two independent fully connected network branches.

[0059] Fault Type Classification Branch: Map the feature vector to a dimensional space equal to the number of fault categories through a fully connected layer, and then process it through a Softmax function to output the predicted probability distribution of each fault category : , , , where and are trainable parameters of the classification branch; is the weight matrix of the classification layer. This is a trainable parameter matrix that maps a feature vector v of dimension D to a space of dimension C (each category corresponds to a score). is the bias vector of the classification layer. A trainable parameter, each category corresponds to a bias value. is the raw output (Logits) of the classification layer. A vector of length C, where each element represents the raw score (not normalized) that the model thinks the input sample belongs to the i-th class. represents the predicted probability of the i-th failure class. Its value is between 0 and 1. represents the classification probability distribution of failure types. A vector of length C, in the form of where denotes the probability that the sample belongs to the i-th class. The global feature v is converted into a probability distribution, clearly representing the model's judgment of failure types and its confidence.

[0060] Failure severity regression branch: Map the feature vector v through another fully connected layer to a one-dimensional space, and use a linear activation function to output the continuous regression estimate of failure severity : , where , is the trainable parameter of the regression branch, is the weight vector of the regression layer. It is a trainable parameter vector that extracts the most relevant information for failure severity from the feature vector v of dimension D. is the global feature vector. It shares the same input feature vector with the classification branch. is the bias term of the regression layer. A trainable scalar value. is the regression estimate of failure severity. A continuous scalar value representing the model's prediction of failure severity. Its specific range and meaning need to be defined before training (for example, 0 represents normal, 1 represents severe failure). The global feature v is converted into a continuous numerical value for quantitative evaluation of the severity of the failure.

[0061] Finally, the model simultaneously outputs the classification probability distribution of failure types and the regression estimate of failure severity . This design allows the model to simultaneously and efficiently complete the two core tasks in fault diagnosis: what is the fault? And how serious is it?

[0062] These four sub-steps constitute the core processing pipeline of the model: first, read comprehensively (BiLSTM extraction) → then focus on labeling (attention focus) → then summarize the essence (feature compression) → finally give qualitative and quantitative conclusions (classification + regression). Just like an experienced engineer first listens to a whole segment of bearing sound, then repeatedly listens to the key abnormal segment, and finally comprehensively judges the failure type and severity.

[0063] ​​S3. The deep learning model is trained using a labeled bearing vibration dataset. The Adam optimizer is used for parameter optimization, with the cross-entropy error function as the loss function. An early stopping mechanism is introduced to prevent overfitting. This optimizes the model parameters and improves generalization ability. The model is trained using labeled data, and the Adam optimizer is used to dynamically adjust the parameters. Cross-entropy error is used as the main loss function, and an early stopping mechanism is combined to prevent overfitting, ensuring stable model convergence and good generalization ability.

[0064] Specifically, in this embodiment, the model training and optimization process in step S3 includes: (1) Design of multi-objective composite loss function: The model training uses a classification-regression joint loss function, with the loss calculated from the cross-entropy error (CEE). and regression loss based on fault impact characteristics constitute: , Among them, the weighting coefficient The training process is dynamically adjusted according to the training cycle, with the initial stage focusing on regression tasks and the later stage focusing on classification tasks. , Let be the weight value in the t-th training epoch. These represent the maximum and minimum values ​​of the weights, respectively. The decay rate. A hyperparameter greater than 0, controlling the weights from... decay to The speed. This is the index for the training epoch. The exponentially decaying term. As t increases, this term gradually approaches 0.

[0065] In the early stages of training (when t is small) The total loss was caused by The primary focus is on enabling the model to initially learn how to assess the severity of faults. As training progresses, Gradually decrease, weight As the size increases, the model begins to focus more on learning the types of faults. This is a curriculum learning strategy.

[0066] The classification loss uses the cross-entropy error function: , Where N is the number of samples and C is the number of fault categories. For the sample In category the true label on the training set, the probability predicted by the model.

[0067] Regression loss combines mean squared error and consistency constraint of fault impact feature: , where, is the true fault degree, is the predicted fault degree, KL is the KL divergence, which measures the difference between two probability distributions, used to ensure that the predicted fault impact feature distribution is consistent with the true distribution . is the regularization strength hyperparameter. Used to control the proportion of the second term KL divergence in the regression loss. is the predicted fault impact feature distribution. The feature distribution derived from the predicted signal or state of the model. is the true fault impact feature distribution. The distribution that the feature (such as kurtosis, pulse indicator) extracted from the true signal obeys.

[0068] The regression loss not only requires the predicted value to be accurate (the first term), but also requires the predicted result to be consistent with the characteristics of the true fault signal in statistical characteristics (the second term), which introduces physical prior knowledge and can make the prediction more reliable.

[0069] (2) Adaptive optimization strategy: use Adam optimizer for parameter optimization, its update rule is: , , , , , where, is the gradient of the loss function with respect to the parameters at the t-th step. is the first moment estimate of the gradient (momentum). Similar to the velocity term in SGD with momentum. is the second moment estimate of the gradient. Used to adaptively adjust the learning rate of each parameter. is the decay rate of the first and second moments. Usually set to 0.9 and 0.999. , is the first and second moment estimates after bias correction. Used to solve the bias caused by initializing to 0. is the updated model parameter at the t-th step. is the global learning rate. is a very small constant to prevent the denominator from being zero.

[0070] The learning rate adopts an exponential decay strategy: , where, is the actual learning rate at the t-th step (or Epoch). is the initial learning rate. is the decay rate. is the number of training steps or Epochs.

[0071] (3) Regularization and Early Stopping Mechanism: The regularization strategy combines Dropout and weight decay, and the Dropout rate is dynamically adjusted with the training progress. The early stopping mechanism is based on the change of the cross-entropy loss on the validation set: Training is stopped when the validation set loss does not decrease for E consecutive periods: , where, is the loss calculated on the validation set at the t-th Epoch. E is the patience value. If the validation set loss does not decrease for E consecutive Epochs, training is stopped.

[0072] (4) Gradient Clipping and Monitoring: Global gradient clipping is used to prevent gradient explosion: , where, is the original gradient vector calculated at the t-th step (or the t-th mini-batch) (which can be the concatenation of all parameters or the gradient of a single layer). L is the 2 norm (i.e., Euclidean distance) of the gradient, which measures the overall size of the gradient. is the set threshold (clip threshold), which is usually an empirical value such as 1.0, 5.0, 10.0, etc.

[0073] The training monitoring system is used to track the change trend of cross-entropy loss and accuracy in real time.

[0074] By carefully designing a dynamically weighted multi-task loss function, an adaptive optimizer, and regularization strategies, this project aims to address the core challenges in bearing fault prediction: Task balancing: Coordinating classification and regression tasks with different properties through dynamic weights. Training stability: Employing the Adam optimizer, learning rate decay, and gradient pruning. Generalization ability: Preventing overfitting through Dropout, weight decay, and early stopping mechanisms. Physical consistency: Introducing a KL divergence term into the regression loss to ensure that the prediction results conform to the physical characteristics of the fault. These strategies collectively ensure that the model can be trained efficiently and stably, producing accurate and reliable prediction results.

[0075] S4. Input the real-time collected vibration data into the trained deep learning model, and simultaneously output the fault type identification result and the fault severity assessment value. The fault severity assessment is based on the fault impact characteristics learned by the model for quantitative calculation. This achieves simultaneous and accurate diagnosis of fault type and severity. The real-time vibration data is input into the trained model, and simultaneously outputs the fault type classification result (e.g., inner ring crack, ball bearing spalling, etc.) and the fault severity quantitative assessment value (e.g., minor, moderate, severe). The severity assessment is based on the fault impact characteristics learned by the model.

[0076] Specifically, in this embodiment, the fault severity assessment in step S4 is based on a quantitative calculation method that fuses multi-scale spectral kurtosis features with nonlinear dynamic parameters. The specific implementation process is as follows: (1) Extracting multi-scale spectral kurtosis features from the preprocessed real-time vibration signal: The signal is decomposed into J scales using maximum overlap discrete wavelet packet transform (MODWPT), and the spectral kurtosis value of each sub-band signal is calculated: , in, For the first The wavelet coefficients of each sub-band are obtained by decomposing the original vibration signal through maximum overlap discrete wavelet packet transform (MODWPT), representing the components of the signal at a specific frequency band j and time t. This indicates a time averaging operation, which calculates the average of the quantity within the parentheses over the entire time axis. This represents the fourth moment of the wavelet coefficients. It reflects the intensity of the impulse component of the signal in that subband. This represents the second moment of the wavelet coefficients. It approximates the power of the signal in that subband.

[0077] Constructing multi-scale spectral kurtosis feature vectors It consists of spectral kurtosis values ​​from J different sub-bands, comprehensively describing the distribution of fault impact characteristics across the entire frequency band. Among them, Represents the center frequency of the j-th sub-band The spectral kurtosis value of the band. It characterizes the degree of prominence of the impulsive fault feature in the band. The higher the value, the stronger the impact component in the band, the greater the possibility of failure.

[0078] (2) Calculate the nonlinear dynamic parameters: Use the phase space reconstruction technique to determine the optimal embedding dimension m and time delay τ by the C-C method, and calculate the correlation dimension and the maximum Lyapunov exponent : , , where is the correlation integral, which measures the proportion of all pairs of points in the phase space whose distance is less than the radius r. It is used to describe the degree of aggregation of the system trajectory. r represents the radius of the hypersphere in the phase space. is the correlation dimension. It is used to quantify the complexity and degree of freedom of the system dynamics. When the system enters a fault state from a normal state, its dynamic characteristics will change, which can be captured. is the distance vector of two adjacent points at the initial time in the phase space; is the distance vector of the two adjacent points after time t; is the norm of the vector, usually referring to the Euclidean distance. is the maximum Lyapunov exponent. It is used to measure the sensitive dependence on initial conditions of the system (i.e., the degree of chaos). Generally means that the system is chaotic, and the size of its value can reflect the degree of disorder of the system, which is related to the development of the fault. The above formula describes the overall changes in the nonlinear dynamics of the bearing-rotor system, which are deep signs of the occurrence and development of the fault.

[0079] (3) Establish a deep feature fusion evaluation model: input the preliminary estimation value output by the regression branch, the multi-scale spectral kurtosis feature K and the nonlinear dynamic parameters to the lightweight feature fusion network: , , where and are the fusion layer parameters, GeLU is the Gaussian error linear unit activation function, and σ is the Sigmoid function that compresses the output to the [0, 1] interval; is the concatenation operation. Concatenate all the above features into a longer comprehensive feature vector. is the fused advanced feature representation. and are the weights and biases of the output layer. Trainable parameters. The final comprehensive failure degree score. A scalar value between 0 (normal) and 1 (severe failure) that quantifies the severity of the failure. The above formula fuses the preliminary prediction of the deep learning model with the signal processing and dynamics-based physical features to obtain a more accurate and reliable comprehensive failure degree score.

[0080] (4) Failure degree grading based on adaptive threshold: dynamically adjust the threshold boundary according to the historical operation data of the equipment, and establish the failure degree grading function:

[0081] where the threshold parameter is adaptively determined from historical data by kernel density estimation: , is the inverse function of the cumulative distribution function of the historical failure degree evaluation value, is the pre-set quantile value. is the cumulative distribution function of the historical data failure degree evaluation value . is the adaptive threshold. It is calculated according to the historical operation data of the equipment itself, and is used to divide different failure levels. This makes the grading standard match the specific situation of the equipment and is more personalized. is the comprehensive failure degree score of the current sample. The final output failure severity level.

[0082] Through multi-source information fusion, such as deep learning prediction: provide end-to-end preliminary estimation. Signal processing features: accurately capture the impact characteristics of the failure. Dynamics system features: reveal the internal state changes of the system. By fusing these three types of information and using adaptive thresholds based on historical data, this method greatly improves the accuracy and reliability of failure degree evaluation, making it not only a mathematical prediction value, but also a diagnostic indicator with clear physical meaning and practical value. The formula can adaptively determine the failure level division threshold according to the historical performance of the equipment itself, achieving a more scientific and personalized health state evaluation.

[0083] S5, generate a diagnostic report containing both failure type and failure degree, when the failure degree evaluation value exceeds the adaptive threshold set based on historical data, trigger a multi-level warning mechanism. Generate structured diagnostic results and trigger intelligent warning. Output a comprehensive report containing both failure type and degree evaluation value, and dynamically trigger a multi-level warning mechanism (such as general warning, advanced warning, emergency warning) based on the adaptive threshold set based on historical data, realizing closed-loop management from diagnosis to decision-making.

[0084] Specifically, in the present embodiment, the diagnosis report generation and early warning mechanism in step S5 are implemented in the following manner: (1) Multi-modal diagnosis report generation: based on fault type classification probability distribution and comprehensive fault severity evaluation value , structured multi-modal diagnosis report is generated, including: fault type confidence analysis: output the maximum probability fault class and its confidence .

[0085] Fault severity quantitative description: according to value, output specific fault severity numerical value.

[0086] Fault evolution trend analysis: based on the value of the last K time windows, calculate the trend index: , where is a sign function, output +1 when the input is positive, -1 when the input is negative, and 0 when the input is zero. K is the size of the time window. represents the number of the last continuous time period (such as the last 10 collection periods) used to calculate the trend. is the fault severity score at the i-th time point; is the fault evolution trend index.

[0087] (2) Adaptive early warning triggering mechanism: build adaptive early warning threshold based on historical operating state , where and are the mean and standard deviation of the value in the historical normal state, respectively is the adjustable sensitivity coefficient; is the adaptive early warning threshold. This threshold is not a fixed value, but is calculated based on the historical health data of the device itself, and has individualization and adaptability. According to the normal operation baseline of the device itself, the early warning threshold is scientifically set to avoid false alarms or omissions caused by individual differences or changes in working conditions when using fixed thresholds.

[0088] Trigger multi-level early warning when any of the following conditions are met: Emergency warning: and ; the fault has been very serious and is worsening, and needs to be handled immediately.

[0089] High-level warning: and ; the fault is serious and is worsening, and needs to be closely monitored and planned for maintenance.

[0090] General warning: and ; the fault just deviates from the normal range and has a worsening trend, prompting a preliminary warning.

[0091] represents that the overall fault is in a worsening trend, represents that the overall fault is in a worsening trend, represents that the fault degree is basically stable. Not only is attention paid to the severity of the current fault, but more importantly, the dynamic development trend of the fault is judged to provide a decision basis for predictive maintenance.

[0092] (3) Maintenance decision support: according to the fault type and severity, specific maintenance suggestions are generated based on a pre-set maintenance knowledge base: , wherein is a rule-based decision function, and the output includes specific maintenance instructions such as continuing monitoring, planned maintenance, and emergency shutdown; this is a pre-defined logical mapping rule, not a trainable parameter. is the most likely fault type, is the current fault degree score. is the fault evolution trend index. is the output maintenance instruction. According to the input fault information, specific suggestions are matched from the pre-defined maintenance knowledge base, for example: (inner ring crack, 0.15, 0.1) -> continue monitoring, pay attention to lubrication; (ball peeling, 0.65, 0.8) -> plan to stop maintenance next week; (retainer broken, 0.92, 0.9) -> emergency shutdown! Replace the bearing immediately. The prediction results of the model are automatically converted into executable maintenance operations, forming a closed loop, and truly realizing the ultimate goal of predictive maintenance.

[0093] The prediction results of the artificial intelligence model are combined with professional knowledge in the operation and maintenance field to realize state perception, intelligent early warning, decision support, and system integration: through standardization (JSON) and real-time communication protocol (MQTT), the diagnosis results are integrated into the existing equipment management system. These formulas and logics together constitute a complete, automated, and intelligent fault prediction and health management (PHM) process.

[0094] (4) Report visualization and push: The diagnosis results are stored in a JSON format structure and displayed in real time on a fault diagnosis dashboard through a Web interface, and warning information is pushed to the mobile terminals of equipment management personnel through the MQTT protocol.

[0095] Embodiment Two As Figure 2As shown, the application provides a fusion attention mechanism and LSTM industrial bearing vibration time series signal fault prediction system architecture diagram, applied to the fusion attention mechanism and LSTM industrial bearing vibration time series signal fault prediction system as described in embodiment one, comprising a signal acquisition and preprocessing module 11, a model construction module 12, a model training module 13, a real-time fault diagnosis module 14, a diagnosis report and early warning module 15.

[0096] The signal acquisition and preprocessing module 11 is used to collect vibration time domain signals during the operation of industrial bearings using vibration sensors, and to preprocess the collected signals, including filtering and noise reduction and normalization processing, to obtain standardized vibration time series data.

[0097] The model construction module 12 is used to collect vibration time domain signals during the operation of industrial bearings using vibration sensors, and to sequentially perform filtering and noise reduction and normalization processing on the collected signals to obtain standardized vibration time series data.

[0098] The model training module 13 is used to train the deep learning model using the labeled bearing vibration data set, optimize the parameters using the Adam optimizer, use the cross-entropy error function as the loss function, and introduce the early stopping mechanism to prevent overfitting.

[0099] The real-time fault diagnosis module 14 is used to input the real-time collected vibration data into the trained deep learning model, and simultaneously output the fault type recognition result and the fault degree evaluation value, wherein the fault degree evaluation is based on the quantitative calculation of the fault impact features learned by the model.

[0100] The diagnosis report and early warning module 15 is used to generate a diagnosis report containing both fault type and fault degree, and when the fault degree evaluation value exceeds the adaptive threshold set based on historical data, a multi-level early warning mechanism is triggered.

[0101] Figure 3 The electronic device provided by an embodiment of the application. As shown in Figure 3 The electronic device at least includes the following parts: a processor 101 and a memory 100, a communication interface 103, and a bus 102.

[0102] In an embodiment of the application, the memory 100 is used to store processor 101 executable instructions, and the processor 101 is configured to execute instructions to implement the method of the first aspect.

[0103] In an embodiment of the application, a computer readable storage medium includes instructions instructing a device to execute the method of the first aspect. For example, the instructions instruct the device to execute the method shown in the flow steps of Figure 1

[0104] ​The program that works in the electronic device according to the embodiment of the present application can be a program that controls a central processing unit (CPU) or the like to realize the functions of the above-described embodiments according to one aspect of the present application (a program that causes a computer to function). Then, the information processed by these systems is temporarily stored in a random access memory (RAM) when it is processed, and then stored in various ROMs such as a read only memory (Flash ROM) and a hard disk drive (HDD), and read out, corrected, and written by the CPU as necessary.

[0105] Note that a part of the electronic device according to the above-described embodiments can also be realized by a computer. In this case, a program for realizing the control function can be recorded in a computer-readable recording medium, and realized by reading the program recorded in the recording medium into a computer and executing it.

[0106] Note that the computer referred to here means a computer built in the electronic device, and a computer including hardware such as an OS and a peripheral device. Further, the computer-readable recording medium means a removable medium such as a floppy disk, a magneto-optical disk, a ROM, a CD-ROM, and a storage system such as a hard disk built in the computer.

[0107] Further, the computer-readable recording medium can include a medium that dynamically stores a program for a short time, such as a communication line in the case of transmitting a program via a network such as the Internet or a communication line such as a telephone line, and a medium that stores a program for a fixed time, such as a volatile memory in a computer that is a server or a client in this case. Further, the above-described program can be a program for realizing a part of the above-described functions, and can also be a program that can realize the above-described functions by being combined with a program already recorded in a computer.

[0108] Further, the electronic device according to the above-described embodiments can also be realized as an aggregate (system group) composed of a plurality of systems. Each system that constitutes the system group can have a part or all of each function or each functional block of the electronic device according to the above-described embodiments. As the system group, all of each function or each functional block of the electronic device can be included.

[0109] Those skilled in the art will recognize that the above embodiments are merely illustrative of the application and should not be taken as limiting the scope of the application, as defined by the following claims. Where necessary, suitable changes and modifications to the above embodiments can be made without departing from the spirit and scope of the application.

Claims

1. A method for predicting faults in industrial bearing vibration timing signals by integrating attention mechanisms and LSTM, characterized in that, Includes the following steps: S1. Use vibration sensors to collect vibration time-domain signals during the operation of industrial bearings. Then, filter, reduce noise, and normalize the collected signals to obtain standardized vibration time-series data. S2. Construct a deep learning model that includes a bidirectional LSTM layer and a coordinate attention mechanism layer. The bidirectional LSTM layer is used to extract bidirectional temporal features from vibration time series data, and the coordinate attention mechanism layer is used to perform time and feature dimension weighting on the hidden state output by the bidirectional LSTM layer to enhance the attention to key fault features. S3. The deep learning model is trained using the labeled bearing vibration dataset. The parameters are optimized using the Adam optimizer, the loss function is the cross-entropy error function, and an early stopping mechanism is introduced to prevent overfitting. S4. Input the real-time collected vibration data into the trained deep learning model, and output the fault type identification result and fault severity assessment value simultaneously. The fault severity assessment is based on the fault impact characteristics learned by the model and is quantitatively calculated. S5. Generate a diagnostic report that includes both the fault type and the fault severity. When the fault severity assessment value exceeds the adaptive threshold set based on historical data, a multi-level early warning mechanism is triggered.

2. The method for predicting faults in industrial bearing vibration timing signals by fusing attention mechanism and LSTM as described in claim 1, characterized in that, The preprocessing in step S1 includes the following sub-steps: S1.1 Signal filtering: An adaptive bandpass filter is used, with a passband frequency range of 0.5 to 3 times the bearing characteristic frequency, in order to retain the fault characteristic frequency components and suppress irrelevant frequency band interference. S1.2 Noise Suppression Processing: A wavelet packet-based threshold denoising algorithm is adopted, which uses the db4 wavelet basis for multi-level wavelet packet decomposition and an adaptive threshold function to perform soft thresholding on the detail coefficients. S1.3 Signal normalization processing: The maximum-minimum normalization method is used to normalize the signal amplitude to a preset numerical range; S1.4 Data Augmentation Processing: Expanding the training dataset using time series data augmentation techniques, including one or more of time warping, amplitude scaling, and adding Gaussian noise; S1.5 Feature extraction preprocessing: Calculate the time-domain feature parameters of the vibration signal, which include one or more of the root mean square value, kurtosis, peak factor and impulse index, and use these features together with the preprocessed time series data as input to the fault feature extraction model.

3. The method for predicting faults in industrial bearing vibration timing signals by fusing attention mechanism and LSTM according to claim 1, characterized in that, The deep learning model constructed in step S2 is a hierarchical feature extraction and fusion network, specifically including: S2.1, Bidirectional Temporal Feature Extraction Layer: A multi-layer bidirectional long short-term memory network BiLSTM is used as the core temporal feature extractor to capture the long-range bidirectional dependencies related to the fault in the vibration signal; S2.2, Coordinate Attention Mechanism Layer: Receives the hidden state sequence of all time steps of the BiLSTM layer, calculates attention weights along the time dimension and feature dimension respectively through the coordinate attention mechanism, generates a two-dimensional attention weight matrix, and recalibrates and weights the output features of BiLSTM. S2.3 Feature Fusion and Compression Layer: The feature sequence weighted by coordinate attention is weighted and summed in the time dimension to compress it into a feature vector of fixed dimension; S2.4 Classification and Regression Output Layer: Receives the fixed-dimensional feature vector, outputs the classification probability of the fault type through a fully connected layer and a Softmax function, and outputs the regression estimate of the fault degree through another fully connected layer and a linear activation function.

4. The method for predicting industrial bearing vibration timing signals by fusing attention mechanism and LSTM according to claim 3, characterized in that, The bidirectional temporal feature extraction layer in step S2.1 satisfies the following conditions: The multilayer bidirectional long short-term memory network BiLSTM has 2 to 4 layers, and the number of hidden units in each layer can be configured between 64 and 256. The BiLSTM layer receives the standardized vibration time series data after preprocessing in step S1. Its forward and backward LSTM units process the input sequence in forward and reverse time order, respectively, and concatenate the hidden states in the two directions of each time step to form a complete hidden state output that integrates bidirectional context information. During training, the BiLSTM network employs weight normalization combined with activation scaling. By reparameterizing the convolutional weight matrix, it satisfies the distribution characteristics of zero mean and unit variance. At the same time, a learnable scaling parameter is introduced before the activation function to stabilize the activation distribution during network training, accelerate convergence, and alleviate the problem of internal covariate bias. The complete hidden state output is not only passed to the subsequent coordinate attention mechanism layer, but its output at the last time step is also used as a global temporal context feature. It is concatenated with the feature vector after attention weighting and summing, and then input into the classification and regression output layer to enhance the model's ability to perceive the overall state of the sequence.

5. The method for predicting industrial bearing vibration timing signals by fusing attention mechanism and LSTM according to claim 4, characterized in that, The implementation process of the coordinate attention mechanism layer in step S2.2 is as follows: Receive the complete hidden state sequence output by the BiLSTM layer, and first perform one-dimensional global pooling operations along the time dimension and feature dimension respectively to generate global context descriptors on the time dimension and feature dimension. The obtained two-dimensional descriptors are concatenated and fused, and then intermediate feature maps are generated through convolutional layers with shared weights and non-linear activation functions. The intermediate feature maps are separated along the original dimension to obtain the temporal attention weight vector and the channel attention weight vector, respectively. The temporal attention weight vector and the channel attention weight vector are multiplied by an outer product to generate a two-dimensional attention weight matrix. The two-dimensional attention weight matrix is ​​multiplied element-wise with the complete hidden state sequence output by BiLSTM to enhance fault-sensitive periods and key feature channels, while suppressing irrelevant information.

6. The method for predicting faults in industrial bearing vibration timing signals by fusing attention mechanism and LSTM according to claim 5, characterized in that, The feature fusion and compression layer in step S2.3 is implemented in the following manner: Receive coordinate attention-weighted feature sequences Where T is the total number of time steps and D is the feature dimension; A temporal fusion method based on attention weights is used to process the weighted feature sequences. Compression is applied along the time dimension to generate a fixed-dimensional global feature vector. : ， in, Indicates the first The weighted feature vector at each time step For the first The fusion weight coefficients for each time step; The fusion weight coefficient The importance scoring function of time step The calculation is performed and normalized using the Softmax function to obtain: ， ， in, This is the weight matrix. and For learnable parameters, It is the hyperbolic tangent activation function; The final global feature vector It integrates the key information most relevant to the fault throughout the time series and uses it as input for subsequent classification and regression output layers.

7. The method for predicting faults in industrial bearing vibration timing signals by fusing attention mechanism and LSTM according to claim 6, characterized in that, The classification and regression output layer in step S2.4 is implemented in the following way: Receive the fixed-dimensional global feature vector This is then simultaneously input into two independent fully connected network branches; Fault type classification branch: The feature vector is processed through a fully connected layer. Mapped to the number of fault categories Using the same dimensional space, and after processing with the Softmax function, the predicted probability distribution for each fault category is output. : ， ， ， in, and These are the trainable parameters for the classification branch; This is the raw output of the classification layer, for each element. This represents the original score by which the model believes the input sample belongs to the i-th category. Represents the predicted probability of the i-th fault category. The classification probability distribution representing the fault type is in the form of: ; Fault severity regression branch: The feature vector is processed through another fully connected layer. Mapped to a one-dimensional space, and using a linear activation function, a continuous regression estimate of the fault severity is output. : ， in, , These are the trainable parameters for the regression branch; Finally, the model synchronously outputs the classification probability distribution of the fault type. Regression estimate of the degree of failure .

8. The method for predicting faults in industrial bearing vibration timing signals by fusing attention mechanism and LSTM according to claim 1, characterized in that, The model training and optimization process in step S3 specifically includes: (1) Multi-objective composite loss function design: The model training adopts a classification-regression joint loss function, which is composed of cross-entropy error (CEE) loss. and regression loss based on fault impact characteristics constitute: ， Among them, the weighting coefficient The training cycle is dynamically adjusted, with the initial stage focusing on regression tasks and the later stage focusing on classification tasks. (2) Adaptive optimization strategy: The Adam optimizer is used for parameter optimization, and the learning rate adopts an exponential decay strategy; (3) Regularization and early stopping mechanism: The regularization strategy combining Dropout and weight decay is adopted. The Dropout rate is dynamically adjusted with the training progress. The early stopping mechanism is based on the change of cross-entropy loss of the validation set. Training is stopped when the validation set loss does not decrease for E consecutive cycles. (4) Gradient clipping and monitoring: Global gradient clipping is used to prevent gradient explosion, and a training monitoring system is used to track the changes in cross-entropy loss and accuracy in real time.

9. The method for predicting faults in industrial bearing vibration timing signals by fusing attention mechanism and LSTM according to claim 1, characterized in that, The fault severity assessment in step S4 is based on a quantitative calculation method that fuses multi-scale spectral kurtosis features with nonlinear dynamic parameters. The specific implementation process is as follows: (1) Extract multi-scale spectral kurtosis features from the preprocessed real-time vibration signal: Decompose the signal into J scales using maximum overlap discrete wavelet packet transform, calculate the spectral kurtosis value of each sub-band signal, and construct a multi-scale spectral kurtosis feature vector K. (2) Calculate nonlinear dynamic parameters: Using phase space reconstruction technology, the optimal embedding dimension m and time delay τ are determined through the CC method, and the correlation dimension is calculated. and the maximum Lyapunov index ; (3) Establish a deep feature fusion evaluation model: the preliminary estimated values ​​output by the regression branch. Multiscale spectral kurtosis characteristics K and nonlinear dynamic parameters The input is fed into a lightweight feature fusion network to obtain the final comprehensive fault severity score. ; (4) Fault severity classification based on adaptive threshold: The threshold boundary is dynamically adjusted according to the historical operating data of the equipment, a fault severity classification function is established, and the classification threshold parameters are adaptively determined from the historical data through the kernel density estimation method.

10. A fault prediction system for industrial bearing vibration timing signals integrating attention mechanism and LSTM, applied to the fault prediction method for industrial bearing vibration timing signals integrating attention mechanism and LSTM as described in any one of claims 1 to 9, characterized in that, The system includes: The signal acquisition and preprocessing module is used to acquire the vibration time-domain signal of the industrial bearing during operation using vibration sensors, and to preprocess the acquired signal, including filtering and noise reduction and normalization, to obtain standardized vibration time series data. The model building module is used to collect vibration time-domain signals of industrial bearings during operation using vibration sensors. The collected signals are then filtered, denoised, and normalized sequentially to obtain standardized vibration time-series data. The model training module is used to train the deep learning model using a labeled bearing vibration dataset. The Adam optimizer is used for parameter optimization, the cross-entropy error function is used as the loss function, and an early stopping mechanism is introduced to prevent overfitting. The real-time fault diagnosis module is used to input the real-time collected vibration data into the trained deep learning model and simultaneously output the fault type identification result and the fault severity assessment value. The fault severity assessment is based on the fault impact characteristics learned by the model for quantitative calculation. The diagnostic report and early warning module is used to generate a diagnostic report that includes both the fault type and the fault severity. When the fault severity assessment value exceeds the adaptive threshold set based on historical data, a multi-level early warning mechanism is triggered.

Images