A mechanical equipment fault prediction method, system, device and storage medium

Abstract

The application discloses a kind of mechanical equipment's fault prediction method, system, device and storage medium, method includes: obtaining the time series data of mechanical equipment;Time series data is input into pre-trained TimesNet model, and target fault prediction result is obtained;Wherein, TimesNet model is introduced in training process New data loss, old data loss, EWC regularization term, knowledge distillation loss and consistency regularization loss are used as comprehensive loss function to carry out model parameter update, and new data loss, old data loss, EWC regularization term, knowledge distillation loss and consistency regularization loss respectively have different weights.The application can improve the fault prediction precision of mechanical equipment.

Description

Technical Field

[0001] This invention belongs to the field of fault prediction, specifically relating to a fault prediction method, system, device, and storage medium for mechanical equipment. Background Technology

[0002] With the rapid development of digitalization in manufacturing, production lines and intelligent equipment have achieved real-time perception of their operating status and environment through sensors, controllers, and smart instruments, generating a large amount of industrial time-series data. Anomaly detection using time-series data can promptly identify abnormal situations, allowing for appropriate measures to prevent accidents.

[0003] As a model designed for time series forecasting, the TimesNet model can capture multiple periodic changes in time series and better simulate complex time series data in the real world. However, existing TimesNet models are prone to catastrophic forgetting during continuous learning, resulting in insufficient prediction accuracy of the trained TimesNet model. Summary of the Invention

[0004] This invention addresses the problems existing in the prior art by providing a method, system, device, and storage medium for predicting faults in mechanical equipment, which can improve the accuracy of fault prediction for mechanical equipment.

[0005] To solve the above technical problems, the present invention provides the following technical solution:

[0006] Construct a multi-source time-series dataset for mechanical equipment. The multi-source time-series data includes one or more of the following: vibration data, temperature data, rotational speed data, pressure data, acoustic data, current data, and historical operation records collected during the operation of the mechanical equipment.

[0007] An anomaly detection model is constructed using the multi-source time-series data to detect anomalies in the operating status of mechanical equipment, and labeled data and pseudo-labeled data are generated based on the anomaly detection results.

[0008] A fault classification model is trained based on the labeled data and the pseudo-labeled data to achieve automatic identification and classification of different fault types.

[0009] A fault prediction model based on TimesNet is constructed, and a comprehensive loss function is constructed by introducing new data loss, old data loss, elastic weight integration EWC regularization term, knowledge distillation loss and consistency regularization loss to achieve parameter optimization of the model during continuous learning.

[0010] The trained fault prediction model is used to analyze real-time collected mechanical equipment operation data to obtain one or more of the following: equipment fault category, fault risk level, fault occurrence probability, or remaining service life prediction results.

[0011] Furthermore, the anomaly detection model is an anomaly detection model constructed based on an autoencoder, variational autoencoder, Transformer, self-supervised learning model, or a combination thereof.

[0012] Furthermore, the fault classification model is a fault classification model constructed based on convolutional neural network (CNN), long short-term memory network (LSTM), Transformer, TimesNet, or a combination thereof.

[0013] Furthermore, the comprehensive loss function includes a new data loss term, an old data loss term, an elastic weight integration EWC regularization term, a knowledge distillation loss term, and a consistency regularization loss term, and each loss term is weighted and summed using preset weights.

[0014] Furthermore, the fault prediction model adopts a continuous learning mechanism, which combines the retention of historical task knowledge with the learning of new task knowledge to achieve continuous optimization and iterative updates of the model in non-stationary industrial environments.

[0015] Furthermore, the method also includes generating equipment health status assessment results and maintenance decision recommendations based on the fault prediction results.

[0016] The present invention also provides a fault prediction system for mechanical equipment used in the aforementioned fault prediction method, comprising:

[0017] A timing data acquisition unit is used to acquire timing data of the mechanical equipment;

[0018] The prediction unit is used to input time-series data into a pre-trained TimesNet model to obtain the target fault prediction result. During the training process, the TimesNet model introduces new data loss, old data loss, EWC regularization term, knowledge distillation loss, and consistency regularization loss as a comprehensive loss function to update the model parameters. The new data loss, the old data loss, the EWC regularization term, the knowledge distillation loss, and the consistency regularization loss each have different weights.

[0019] The present invention also provides a control device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the fault prediction method for mechanical equipment as described in any one of the present invention.

[0020] The present invention also provides a computer-readable storage medium storing computer-executable instructions for executing the fault prediction method for mechanical equipment according to any one of the present invention.

[0021] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:

[0022] This invention uses the TimesNet model to predict mechanical equipment faults. During the training process, the TimesNet model incorporates new data loss, old data loss, EWC regularization term, knowledge distillation loss, and consistency regularization loss as a comprehensive loss function to update model parameters. This prevents catastrophic forgetting of the model and enables dynamic updating of model parameters, thereby improving the accuracy of the TimesNet model in predicting mechanical equipment faults.

[0023] A control device according to a third aspect embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the fault prediction method for mechanical equipment as described in the first aspect embodiment above. Since the control device employs all the technical solutions of the fault prediction method for mechanical equipment described in the above embodiments, it possesses at least all the beneficial effects brought about by the technical solutions of the above embodiments.

[0024] According to a fourth aspect embodiment of the present invention, a computer-readable storage medium stores computer-executable instructions for performing the fault prediction method for mechanical equipment as described in the first aspect embodiment above. Since the computer-readable storage medium employs all the technical solutions of the fault prediction method for mechanical equipment described in the above embodiments, it possesses at least all the beneficial effects brought about by the technical solutions of the above embodiments.

[0025] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0026] Figure 1 This is a flowchart of a fault prediction method for mechanical equipment according to an embodiment of the present invention. Detailed Implementation

[0027] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0028] In the description of this invention, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance, or implicitly indicating the number of technical features indicated, or implicitly indicating the order of the technical features indicated.

[0029] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the drawings and are only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0030] In the description of this invention, it should be noted that, unless otherwise explicitly defined, terms such as "setting," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0031] The following will combine Figure 1 The method for predicting the failure of mechanical equipment according to the embodiments of the present invention will be clearly and completely described. Obviously, the embodiments described below are some embodiments of the present invention, not all embodiments.

[0032] refer to Figure 1 , Figure 1 This is a flowchart of a fault prediction method for mechanical equipment according to an embodiment of the present invention. This embodiment provides a fault prediction method for mechanical equipment, with the following steps:

[0033] S1. Construct a time-series dataset for mechanical equipment, including the mechanical equipment's runtime sequence data, as well as the corresponding anomaly detection results and fault classification detection results; the mechanical equipment's runtime sequence data includes, but is not limited to, the mechanical equipment's vibration data, temperature data, rotational speed data, ultrasonic data, and pressure data; the time-series data is detected by corresponding multi-dimensional sensors, and the real-time time-series data is transmitted to the data center via wireless transmission for fault prediction.

[0034] S2. Construct a time-series dataset of mechanical equipment fault detection neural network based on TimesNet. Train the neural network using the time series dataset of mechanical equipment to obtain a fault prediction model for mechanical equipment. During the training process, introduce new data loss, old data loss, EWC regularization term, knowledge distillation loss and consistency regularization loss as a comprehensive loss function to update the model parameters. The new data loss, the old data loss, the EWC regularization term, the knowledge distillation loss and the consistency regularization loss have different weights.

[0035] S3. Input the real-time collected mechanical equipment operation data into the mechanical equipment fault prediction model to obtain the target fault prediction result.

[0036] The TimesNet model in this invention is an improved TimesNet model that integrates Convolutional Neural Networks (CNNs) and Transformers. First, it utilizes upscaling convolution techniques to map one-dimensional time-series data to a higher-dimensional space. Specifically, the one-dimensional time-series data acquired by multi-dimensional sensors is divided into fixed-length time windows to form a two-dimensional time-feature matrix. To accommodate the processing of the convolutional neural network, channel dimensions are added, transforming the two-dimensional time-feature matrix into a three-dimensional structure.

[0037] Building upon this, multiple one-dimensional or two-dimensional convolutional layers are designed. The convolutional kernels slide along the time and feature dimensions to extract local spatiotemporal features. The kernel size and stride can be adjusted according to the data characteristics; for example, using a kernel size of... The convolutional kernel captures patterns in the time dimension, or The convolutional kernels capture patterns in the feature dimension. The output of the convolutional layer passes through a non-linear activation function (such as ReLU) and a pooling layer to generate a high-level feature representation.

[0038] Next, the features output by the convolutional layer are flattened or reduced in dimensionality to obtain a series of feature vectors. Where L is the sequence length. To preserve the information of the time series, positional encoding is added, embedding the temporal position information into the feature vector:

[0039] ;

[0040] in, Encode the location vectors. Then, use these feature vectors. The input is fed into the Transformer encoder. The Transformer's self-attention mechanism captures global dependencies over long time spans. The Transformer encoder encodes the input features using multi-head self-attention and a feedforward neural network to generate multi-step predictions of the device's state—predicted sensor readings (such as vibration, temperature, and rotational speed) over multiple future time steps. These predictions (i.e., target fault predictions) reflect the device's operating status and help detect anomalies or impending faults.

[0041] It should be noted that the basic structure and principles of Convolutional Neural Networks (CNN) and Transformers are existing technologies known to those skilled in the art, and will not be elaborated upon here.

[0042] To prevent catastrophic forgetting in the TimesNet model during continuous learning, this invention proposes a hybrid continuous learning strategy combining exponentially weighted average (EWA), elastic weight integration (EWC), knowledge distillation, and semi-supervised learning. During the dynamic update of model parameters, a comprehensive loss function is defined as follows:

[0043] ;

[0044] in, For new data loss, Due to the loss of old data, The weight coefficients range from [0, 1]. EWA assigns higher weights to recent data, enabling the model to quickly adapt to changes in device status.

[0045] For EWC regularization terms, The balancing coefficient is a hyperparameter used to balance the effects of various terms in the loss function. Its optimal value is determined by hyperparameter tuning on the validation set or by using methods such as Bayesian optimization. For parameters The importance of this parameter, estimated from the Fisher information matrix, indicates its significance to model performance in previous tasks. This represents the parameters of the current model (such as the weights and biases of a neural network). This represents the parameters of the old model. EWC is used to add a regularization term to the loss function, penalizing the updates of important parameters, maintaining their stability, and preventing the model from forgetting important knowledge when learning new data.

[0046] For knowledge distillation loss, KL , and These are the output probability distributions of the old model and the new model, respectively. The balancing coefficient is a hyperparameter used to balance the knowledge distillation loss with other terms in the total loss. Its value is determined through hyperparameter tuning on the validation set. Knowledge distillation uses the old model as the teacher model and the new model as the student model, minimizing the difference between their outputs while preserving the knowledge from the old model.

[0047] For consistency regularization loss, , and The model corresponds to the original data. and enhanced data The output, The tradeoff coefficient is a hyperparameter used to balance the consistency regularization loss with other terms in the total loss. Its value is determined by hyperparameter tuning on the validation set.

[0048] By minimizing the comprehensive loss function The Adam optimization algorithm is used to update the model parameters, thereby achieving dynamic model updates.

[0049] This invention, through the introduction of new data loss and old data loss, and by assigning higher weights to recent data via EWA, can prevent catastrophic forgetting in the model. The introduction of EWC regularization effectively prevents overfitting and improves the model's generalization ability. The introduction of knowledge distillation loss compresses large models into smaller ones, significantly reducing computational cost and memory usage. The distilled student model typically retains most of the accuracy of the teacher model while being more computationally efficient, making it suitable for deployment in resource-constrained environments. Furthermore, the knowledge in the large model helps the smaller model better learn patterns and rules in the data, thereby improving its generalization ability. The introduction of consistency regularization loss mitigates overfitting (consistency regularization, by introducing consistency constraints on model output, alleviates problems that may arise from discriminative losses (such as cross-entropy loss), which helps avoid the model being affected by noise in the training data). Consistency regularization can improve a model's generalization ability by: 1) overfitting to minor changes, thereby enhancing generalization ability (consistency regularization encourages the model to respond consistently to small perturbations in the input, meaning the model can maintain stable outputs when faced with local changes in the data. This mechanism forces the model to learn robust representations of the data, rather than relying solely on global information, thus improving the model's predictive performance on new data); 2) mitigate the impact of label noise (when label noise is present, consistency regularization can reduce the impact of label noise by encouraging the model to respond consistently to different variations of unlabeled data, reducing the likelihood of assigning unlabeled data to the wrong category); and 3) enable multi-view learning (consistency regularization methods typically require multiple transformations of the same sample to generate different versions of the input, allowing the model to observe the data from multiple perspectives, learn richer feature representations, and help better distinguish different categories).

[0050] It should be noted that the specific principles of exponential weighted average (EWA), elastic weight integration (EWC), knowledge distillation, and semi-supervised learning are existing technologies known to those skilled in the art, and will not be elaborated here.

[0051] According to the mechanical equipment fault prediction method of the present invention, the present invention uses the TimesNet model to predict mechanical equipment faults. During the training process of the TimesNet model, new data loss, old data loss, EWC regularization term, knowledge distillation loss and consistency regularization loss are introduced as a comprehensive loss function to update the model parameters, which prevents catastrophic forgetting of the model and realizes dynamic updating of model parameters, thereby improving the fault prediction accuracy of the TimesNet model for mechanical equipment.

[0052] In some embodiments of the present invention, the TimesNet model includes an anomaly detection model and a fault classification model. Time-series data is input into a pre-trained TimesNet model to obtain target fault prediction results, including:

[0053] Time series data is input into the anomaly detection model to obtain anomaly detection results;

[0054] Time series data is input into the classification model to obtain fault classification and detection results;

[0055] The target fault prediction result is determined based on the anomaly detection result and the fault classification detection result.

[0056] During the training process, both the anomaly detection model and the classification model introduce new data loss, old data loss, EWC regularization term, knowledge distillation loss, and consistency regularization loss as a comprehensive loss function to update the model parameters. The new data loss, old data loss, EWC regularization term, knowledge distillation loss, and consistency regularization loss have different weights.

[0057] The TimesNet model has multiple tasks, including long-term and short-term prediction, missing value imputation, anomaly detection, and classification. In the final prediction, the fusion of anomaly detection results and fault classification detection results enables the final fault prediction of mechanical equipment.

[0058] Preferably, the target fault prediction result is determined based on the anomaly detection result and the fault classification detection result, including:

[0059] Obtain the first actual detection data corresponding to the anomaly detection result, and the second actual detection data corresponding to the fault classification detection result;

[0060] The reconstruction error is calculated based on the anomaly detection results and the first actual detection data.

[0061] The prediction accuracy is calculated based on the fault classification detection results and the second actual detection data.

[0062] The first weight parameter corresponding to the anomaly detection result and the second weight parameter corresponding to the fault classification detection result are determined based on the reconstruction error and the prediction accuracy.

[0063] The target fault prediction result is determined based on the anomaly detection result, the first weight parameter, the fault classification detection result, and the second weight parameter.

[0064] The constraint formula for reconstruction error is:

[0065] ;

[0066] in, This is an abnormal detection result. This is the first actual test data.

[0067] The constraint formula for the first weight parameter is:

[0068] ;

[0069] The constraint formula for the second weight parameter is:

[0070] ;

[0071] in, To predict accuracy.

[0072] The constraint formula for the target fault prediction result is:

[0073] ;

[0074] in, For the fault classification detection results, This is an abnormal detection result.

[0075] It should be noted that the calculation principle of prediction accuracy is existing technology known to those skilled in the art, and will not be elaborated here.

[0076] During the prediction process, anomaly detection results and fault classification detection results are integrated, and an adaptive weight adjustment mechanism is adopted. The weights of the two are dynamically adjusted according to the reconstruction error and prediction accuracy, so that as the classification model is gradually improved, their weights in the prediction are gradually increased, thereby improving the overall prediction accuracy.

[0077] In some embodiments of the present invention, the classification model is trained through the following steps:

[0078] Adaptive labeling is performed on anomaly detection results corresponding to reconstruction errors greater than the first threshold to obtain labeled data;

[0079] Pseudo-labels are generated using labeled data and anomaly detection results corresponding to reconstruction errors greater than the second threshold and less than the first threshold; the second threshold represents the lower limit value for triggering fault warnings.

[0080] The classification model is trained by using both labeled and pseudo-labeled data as training data.

[0081] Understandably, a fault warning is triggered when the reconstruction error exceeds the second threshold. For reconstruction errors that are large (i.e., greater than the first threshold), an improved adaptive labeling mechanism, combined with an anomaly pattern recognition algorithm, is used to automatically identify and label faults or specific fault types based on anomaly detection results through pattern analysis of the reconstruction error, thereby obtaining labeled data and improving labeling efficiency and accuracy. For data with small errors but still exceeding the second threshold (i.e., anomaly detection results corresponding to reconstruction errors greater than the second threshold but less than the first threshold), a semi-supervised learning strategy is adopted. Through methods such as consistency regularization and pseudo-label generation, pseudo-label data is automatically generated to enrich the training data and improve the model's generalization ability.

[0082] The classification model is trained using both labeled and pseudo-labeled data. During training, labeled data includes previously labeled samples as well as newly labeled samples through an adaptive labeling mechanism. Pseudo-labeled data consists of pseudo-labels of unlabeled data generated through a semi-supervised learning strategy. Both are considered as new training data and are used together with the knowledge from the old model to train the new model. A hybrid continuous learning strategy is continued to be used during training to prevent catastrophic forgetting of the model.

[0083] In some embodiments of the present invention, pseudo-labels are generated using labeled data and anomaly detection results corresponding to reconstruction errors greater than a second threshold and less than a first threshold, resulting in pseudo-label data, including:

[0084] An initial model for the classification model is obtained by training labeled data;

[0085] The anomaly detection results corresponding to reconstruction errors greater than the second threshold and less than the first threshold are input into the initial model to predict pseudo-labeled data.

[0086] It is understandable that the principle of pseudo-label generation is as follows: a preliminary model trained with labeled data is used to predict unlabeled data and generate labels for this unlabeled data. These predicted labels are pseudo-labels.

[0087] In some embodiments, samples whose predicted probability by the initial model is higher than a set threshold can be selected as pseudo-labels and added to the training set as pseudo-label data.

[0088] In some embodiments of the present invention, anomaly detection results corresponding to reconstruction errors greater than a second threshold and less than a first threshold are input into the initial model to predict pseudo-labeled data, including:

[0089] For anomaly detection results whose reconstruction error is greater than the second threshold but less than the first threshold, random noise, time offset and data scaling are added sequentially to generate enhanced sample data.

[0090] For anomaly detection results with reconstruction errors greater than the second threshold and less than the first threshold, and for sample data, consistency regularization is performed to obtain unlabeled data.

[0091] Inputting unlabeled data into the initial model yields pseudo-labeled data.

[0092] It is understandable that enhancement processing and consistency regularization processing can keep the model's output consistent and enhance the model's utilization of unlabeled data. The specific principles are existing technologies known to those skilled in the art and will not be elaborated here.

[0093] In some embodiments of the present invention, the method further includes:

[0094] Obtain the operating time, purchase time, maintenance records, and number of abnormalities of the mechanical equipment;

[0095] By inputting the running time, purchase time, maintenance records, and number of anomalies into the pre-trained remaining useful life prediction model, the failure risk prediction results and remaining useful life prediction results are obtained.

[0096] Understandably, by inputting operating time, purchase time, maintenance records, and the number of anomalies into a pre-trained remaining useful life prediction model, the resulting failure risk prediction and remaining useful life prediction are obtained. This is achieved using a multi-task learning method, which simultaneously predicts the equipment's failure risk and remaining useful life curves. The multi-task learning method employs a shared network structure, and the loss function is:

[0097] ;

[0098] in, To predict losses from failures, Predicting losses for remaining lifetime, This is a weighting factor.

[0099] By sharing feature representations, the accuracy of remaining lifetime prediction can be improved, the operational reliability of equipment can be increased, and maintenance costs can be reduced.

[0100] According to a second aspect of the present invention, a fault prediction system for mechanical equipment includes a time-series data acquisition unit and a prediction unit.

[0101] The timing data acquisition unit is used to acquire timing data of mechanical equipment;

[0102] The prediction unit is used to input time-series data into a pre-trained TimesNet model to obtain the target fault prediction result. During the training process, the TimesNet model introduces new data loss, old data loss, EWC regularization term, knowledge distillation loss and consistency regularization loss as a comprehensive loss function to update the model parameters. New data loss, old data loss, EWC regularization term, knowledge distillation loss and consistency regularization loss have different weights.

[0103] It is understandable that time-series data includes, but is not limited to, vibration data, temperature data, rotational speed data, ultrasonic data, and pressure data of mechanical equipment, which are detected by corresponding multi-dimensional sensors and transmitted to the data center in real time via wireless transmission for fault prediction.

[0104] The TimesNet model in this invention is an improved TimesNet model that integrates Convolutional Neural Networks (CNNs) and Transformers. First, it utilizes upscaling convolution techniques to map one-dimensional time-series data to a higher-dimensional space. Specifically, the one-dimensional time-series data acquired by multi-dimensional sensors is divided into fixed-length time windows to form a two-dimensional time-feature matrix. To accommodate the processing of the convolutional neural network, channel dimensions are added, transforming the two-dimensional time-feature matrix into a three-dimensional structure.

[0105] Building upon this, multiple one-dimensional or two-dimensional convolutional layers are designed. The convolutional kernels slide along the time and feature dimensions to extract local spatiotemporal features. The kernel size and stride can be adjusted according to the data characteristics; for example, using a kernel size of... The convolutional kernel captures patterns in the time dimension, or The convolutional kernels capture patterns in the feature dimension. The output of the convolutional layer passes through a non-linear activation function (such as ReLU) and a pooling layer to generate a high-level feature representation.

[0106] Next, the features output by the convolutional layer are flattened or reduced in dimensionality to obtain a series of feature vectors. Where L is the sequence length. To preserve the information of the time series, positional encoding is added, embedding the temporal position information into the feature vector:

[0107] ; Formula (1)

[0108] in, Encode the location vectors. Then, use these feature vectors. The input is fed into the Transformer encoder. The Transformer's self-attention mechanism captures global dependencies over long time spans. The Transformer encoder encodes the input features using multi-head self-attention and a feedforward neural network to generate multi-step predictions of the device's state—predicted sensor readings (such as vibration, temperature, and rotational speed) over multiple future time steps. These predictions (i.e., target fault predictions) reflect the device's operating status and help detect anomalies or impending faults.

[0109] It should be noted that the basic structure and principles of Convolutional Neural Networks (CNN) and Transformers are existing technologies known to those skilled in the art, and will not be elaborated upon here.

[0110] To prevent catastrophic forgetting in the TimesNet model during continuous learning, this invention proposes a hybrid continuous learning strategy combining exponentially weighted average (EWA), elastic weight integration (EWC), knowledge distillation, and semi-supervised learning. During the dynamic update of model parameters, a comprehensive loss function is defined as follows:

[0111] ; Formula (2)

[0112] in, For new data loss, Due to the loss of old data, The weight coefficients range from [0, 1]. EWA assigns higher weights to recent data, enabling the model to quickly adapt to changes in device status.

[0113] For EWC regularization terms, The balancing coefficient is a hyperparameter used to balance the effects of various terms in the loss function. Its optimal value is determined by hyperparameter tuning on the validation set or by using methods such as Bayesian optimization. For parameters The importance of this parameter, estimated from the Fisher information matrix, indicates its significance to model performance in previous tasks. This represents the parameters of the current model (such as the weights and biases of a neural network). This represents the parameters of the old model. EWC is used to add a regularization term to the loss function, penalizing the updates of important parameters, maintaining their stability, and preventing the model from forgetting important knowledge when learning new data.

[0114] For knowledge distillation loss, KL , and These are the output probability distributions of the old model and the new model, respectively. The balancing coefficient is a hyperparameter used to balance the knowledge distillation loss with other terms in the total loss. Its value is determined through hyperparameter tuning on the validation set. Knowledge distillation uses the old model as the teacher model and the new model as the student model, minimizing the difference between their outputs while preserving the knowledge from the old model.

[0115] For consistency regularization loss, , and The model corresponds to the original data. and enhanced data The output, The tradeoff coefficient is a hyperparameter used to balance the consistency regularization loss with other terms in the total loss. Its value is determined by hyperparameter tuning on the validation set.

[0116] By minimizing the comprehensive loss function The Adam optimization algorithm is used to update the model parameters, thereby achieving dynamic model updates.

[0117] This invention, through the introduction of new data loss and old data loss, and by assigning higher weights to recent data via EWA, can prevent catastrophic forgetting in the model. The introduction of EWC regularization effectively prevents overfitting and improves the model's generalization ability. The introduction of knowledge distillation loss compresses large models into smaller ones, significantly reducing computational cost and memory usage. The distilled student model typically retains most of the accuracy of the teacher model while being more computationally efficient, making it suitable for deployment in resource-constrained environments. Furthermore, the knowledge in the large model helps the smaller model better learn patterns and rules in the data, thereby improving its generalization ability. The introduction of consistency regularization loss mitigates overfitting (consistency regularization, by introducing consistency constraints on model output, alleviates problems that may arise from discriminative losses (such as cross-entropy loss), which helps avoid the model being affected by noise in the training data). Consistency regularization can improve a model's generalization ability by: 1) overfitting to minor changes, thereby enhancing generalization ability (consistency regularization encourages the model to respond consistently to small perturbations in the input, meaning the model can maintain stable outputs when faced with local changes in the data. This mechanism forces the model to learn robust representations of the data, rather than relying solely on global information, thus improving the model's predictive performance on new data); 2) mitigate the impact of label noise (when label noise is present, consistency regularization can reduce the impact of label noise by encouraging the model to respond consistently to different variations of unlabeled data, reducing the likelihood of assigning unlabeled data to the wrong category); and 3) enable multi-view learning (consistency regularization methods typically require multiple transformations of the same sample to generate different versions of the input, allowing the model to observe the data from multiple perspectives, learn richer feature representations, and help better distinguish different categories).

[0118] It should be noted that the specific principles of exponential weighted average (EWA), elastic weight integration (EWC), knowledge distillation, and semi-supervised learning are existing technologies known to those skilled in the art, and will not be elaborated here.

[0119] According to the mechanical equipment fault prediction system of the present invention, the present invention uses the TimesNet model to predict mechanical equipment faults. During the training process of the TimesNet model, new data loss, old data loss, EWC regularization term, knowledge distillation loss and consistency regularization loss are introduced as a comprehensive loss function to update the model parameters, which prevents catastrophic forgetting of the model and realizes dynamic updating of model parameters, thereby improving the fault prediction accuracy of the TimesNet model for mechanical equipment.

[0120] Since the mechanical equipment fault prediction system adopts all the technical solutions of the mechanical equipment fault prediction method in the above embodiments, it has at least all the beneficial effects brought about by the technical solutions in the above embodiments, and will not be repeated here.

[0121] In addition, one embodiment of the present invention provides a control device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor and the memory can be connected via a bus or other means.

[0122] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0123] The non-transient software program and instructions required to implement the mechanical equipment fault prediction method of the above embodiments are stored in memory. When executed by the processor, the mechanical equipment fault prediction method of the above embodiments is executed.

[0124] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0125] Furthermore, one embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions that are executed by a processor or controller, for example, by the processor of the above embodiment, such that the processor performs the fault prediction method for mechanical equipment in the above embodiment.

[0126] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0127] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for predicting faults in mechanical equipment, characterized in that, The steps are as follows: S1. Construct a multi-source heterogeneous time-series dataset for mechanical equipment. The multi-source heterogeneous time-series data includes one or more of the following collected during equipment operation: vibration data, temperature data, current data, acoustic data, speed data, pressure data, equipment operation logs, maintenance records, and historical fault records. S2. Preprocess and extract features from the multi-source heterogeneous time-series data to obtain time-domain features, frequency-domain features, time-frequency-domain features, and equipment operating status features; wherein, the frequency-domain features include one or more of the following: Fast Fourier Transform (FFT) features, envelope spectrum features, and kurtosis features; S3. Construct a fault prediction neural network based on TimesNet multi-source heterogeneous data fusion. Train the fault prediction neural network using the multi-source heterogeneous time series dataset to obtain a mechanical equipment fault prediction model. During the training process, introduce new data loss, old data loss, elastic weight integration EWC regularization term, knowledge distillation loss and consistency regularization loss to construct a comprehensive loss function. Continuously learn and update the model parameters to achieve dynamic adaptation to non-stationary industrial environments. S4. Input the real-time collected mechanical equipment operation data into the fault prediction model to obtain equipment anomaly detection results, fault classification results, and fault risk prediction results; S5. Based on the anomaly detection results, fault classification results, and fault risk prediction results, generate fault diagnosis explanation information; the fault diagnosis explanation information includes one or more of the following: fault type, fault probability, contribution of key sensors, contribution of key features, and fault influencing factors. S6. Generate equipment health status assessment results and maintenance decision suggestions based on the fault diagnosis interpretation information, and output target fault prediction results; S7. Upload the model parameters, equipment operation data and prediction results to the remote management platform, and update the fault prediction model online through the remote upgrade mechanism to achieve continuous model optimization and iteration.

2. The fault prediction method for mechanical equipment according to claim 1, characterized in that, The TimeNet-based mechanical equipment fault detection neural network includes an anomaly detection network and a fault classification network. The neural network is trained using a time-series dataset of mechanical equipment, including: Using mechanical equipment runtime sequence data as input and corresponding anomaly detection results as output, an anomaly detection network is trained to obtain an anomaly detection model. A fault classification network is trained by using the mechanical equipment's runtime sequence data as input and the corresponding fault classification and detection results as output to obtain a fault classification model. The target fault prediction result is determined based on the anomaly detection result and the fault classification detection result.

3. The fault prediction method for mechanical equipment according to claim 2, characterized in that, The target fault prediction result is determined based on the anomaly detection result and the fault classification detection result, including: Obtain the first actual detection data corresponding to the anomaly detection result, and the second actual detection data corresponding to the fault classification detection result; The reconstruction error is calculated based on the anomaly detection results and the first actual detection data. The prediction accuracy is calculated based on the fault classification detection results and the second actual detection data. The first weight parameter corresponding to the anomaly detection result and the second weight parameter corresponding to the fault classification detection result are determined based on the reconstruction error and the prediction accuracy. The target fault prediction result is determined based on the anomaly detection result, the first weight parameter, the fault classification detection result, and the second weight parameter.

4. The fault prediction method for mechanical equipment according to claim 3, characterized in that, Adaptive labeling is performed on the anomaly detection results corresponding to reconstruction errors greater than a first threshold to obtain labeled data; Using the labeled data and the anomaly detection results corresponding to reconstruction errors greater than the second threshold and less than the first threshold, pseudo-labels are generated to obtain pseudo-label data; The second threshold represents the lower limit value for triggering a fault warning; The fault classification model is trained by using labeled data and pseudo-labeled data together as training data; wherein, the fault classification model is used to identify the corresponding fault type based on the mechanical equipment operation data and output the probability value corresponding to each fault type; the fault type includes one or more of the following: bearing fault, gear fault, rotor imbalance, rotor misalignment, mechanical loosening fault, lubrication abnormality fault, and other equipment faults.

5. The fault prediction method for mechanical equipment according to claim 4, characterized in that, Using the labeled data and the anomaly detection results corresponding to reconstruction errors greater than the second threshold and less than the first threshold, pseudo-labels are generated to obtain pseudo-label data, including: The initial model of the fault classification model is obtained by training the labeled data. The anomaly detection result corresponding to the reconstruction error being greater than the second threshold and less than the first threshold is input into the initial model to obtain the predicted probability corresponding to each fault category. When the maximum predicted probability is greater than the preset confidence threshold, the corresponding fault category is assigned as a pseudo-label to the anomaly detection result, generating pseudo-label data; The fault classification model is updated by using the pseudo-labeled data and the labeled data together, thereby achieving semi-supervised learning training of the fault classification model.

6. The fault prediction method for mechanical equipment according to claim 5, characterized in that, The anomaly detection results corresponding to reconstruction errors greater than the second threshold and less than the first threshold are input into the initial model to predict pseudo-labeled data, including: For the anomaly detection results corresponding to reconstruction errors greater than the second threshold and less than the first threshold, random noise, time offset, and data scaling are added sequentially to generate enhanced sample data. For the anomaly detection results corresponding to reconstruction errors greater than the second threshold and less than the first threshold, and the sample data, consistency regularization processing is performed to obtain unlabeled data; Inputting unlabeled data into the initial model yields pseudo-labeled data.

7. The fault prediction method for mechanical equipment according to claim 1, characterized in that, The constraint formula for reconstruction error is: ； in, This is an abnormal detection result. This is the first actual test data; The constraint formula for the first weight parameter is: ； The constraint formula for the second weight parameter is: ； in, To improve prediction accuracy; The constraint formula for the target fault prediction result is: ； in, For the fault classification detection results, This is an abnormal detection result.

8. A fault prediction system for mechanical equipment used in the fault prediction method according to any one of claims 1-7, characterized in that, include: A timing data acquisition unit is used to acquire timing data of the mechanical equipment; The prediction unit is used to input time-series data into a pre-trained TimesNet model to obtain the target fault prediction result. During the training process, the TimesNet model introduces new data loss, old data loss, EWC regularization term, knowledge distillation loss, and consistency regularization loss as a comprehensive loss function to update the model parameters. The new data loss, the old data loss, the EWC regularization term, the knowledge distillation loss, and the consistency regularization loss each have different weights.

9. A control device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the fault prediction method for mechanical equipment as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are used to execute the fault prediction method for mechanical equipment as described in any one of claims 1 to 7.

Images