An automated equipment failure prediction method and system based on big data analysis
Patent Information
- Application Number
- CN202611067434.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-17
- Publication Date
- 2026-10-30
AI Technical Summary
随着工业4.0和智能制造的推进，自动化仪表的数量和复杂度持续增长，传统的定期维护和事后维修模式已难以满足高可靠性要求，基于状态的预测性维护成为行业发展趋势
[0074]1.本发明通过小波包分解提取信号的时频联合能量特征与能量熵，并引入多头注意力机制对多源异构特征进行自适应加权融合，相较于传统单一特征提取方法，融合后的特征向量能够更全面地刻画仪表健康状态，使得故障识别准确率提升。
Smart Images
Figure CN123012399A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial automation and intelligent operation and maintenance technology, specifically a method and system for predicting automated equipment faults based on big data analysis. Background Technology
[0002] Automated instruments are core equipment for parameter detection, control, and regulation in industrial production processes, and their operational reliability directly affects production safety and efficiency. With the advancement of Industry 4.0 and intelligent manufacturing, the number and complexity of automated instruments continue to grow. Traditional periodic maintenance and reactive repair methods are no longer sufficient to meet high reliability requirements, and condition-based predictive maintenance has become an industry trend.
[0003] Existing methods for predicting faults in automated instruments mostly employ single time-domain statistical features or simple frequency-domain analysis, failing to fully exploit the joint time-frequency features in instrument operating data. While wavelet packet decomposition has been used in some studies, it is typically applied only to single signal processing and lacks an effective fusion mechanism for multi-source heterogeneous data, resulting in low feature dimensionality and significant information loss.
[0004] Traditional methods often employ rule-based judgments with fixed thresholds or single machine learning models, which cannot effectively capture the long-term dependencies and nonlinear dynamic changes in the instrument degradation process. While time series models such as LSTM can process sequential data, standard LSTM has limited ability to focus on key time steps, and single models lack generalization ability when facing different failure modes.
[0005] Existing early warning systems mostly employ fixed thresholds or static judgments based on the 3σ principle, failing to consider instrument operating condition drift and data distribution changes. This leads to frequent false alarms or missed alarms during operating condition switching, severely impacting maintenance personnel's trust and response efficiency. Existing methods are largely black-box predictions, unable to explain the basis for fault predictions, and their fixed data acquisition strategies fail to dynamically adjust acquisition resources based on feature importance, resulting in wasted storage and computing resources.
[0006] Therefore, there is an urgent need for an automated instrument fault prediction method and system that can fully utilize multi-source big data, possess strong feature extraction and fusion capabilities, adopt improved prediction models, and support dynamic early warning and resource adaptive optimization.
[0007] The information disclosed above in this background section is only for enhancing the understanding of the background section of this invention, and therefore may include prior art that is not known to those skilled in the art. Summary of the Invention
[0008] The purpose of this invention is to provide an automated equipment fault prediction method and system based on big data analysis to solve the problems in the prior art.
[0009] To achieve the above objectives, the present invention provides the following technical solution: an automated equipment fault prediction method based on big data analysis, comprising the following steps:
[0010] S1: Real-time acquisition of multi-source operating data from automated instruments, including process variable data, environmental variable data, and instrument status data;
[0011] S2: Preprocess the multi-source operating data, including missing value imputation, outlier detection, and data normalization;
[0012] S3: Perform multi-dimensional feature extraction and fusion based on the preprocessed data to obtain a comprehensive feature vector;
[0013] S4: Input the comprehensive feature vector into the improved fault prediction model and output the fault probability and remaining service life prediction values;
[0014] S5: Generate graded early warning information based on the failure probability and remaining service life prediction.
[0015] Preferably, the outlier detection in step S2 employs an improved Z-score algorithm based on a sliding window, specifically including:
[0016] Set the sliding window length to W, for the data at the i-th time point i Calculate local statistics within the window:
[0017]
[0018]
[0019] The improved Z-score is as follows:
[0020]
[0021] Where k is the adaptive coefficient. The coefficient of variation within the window. To prevent division by zero constant; when >Z th If a value is identified as an outlier, it is filled using weighted interpolation of adjacent valid values. The filling formula is as follows:
[0022]
[0023] in Let i be the set of valid samples surrounding anomaly point i. This is the attenuation coefficient.
[0024] Preferably, the multidimensional feature extraction in step S3 employs wavelet packet decomposition to analyze the signal. Perform J-level wavelet packet decomposition:
[0025]
[0026] in Let be the wavelet packet coefficients of the k-th node in the j-th layer. The wavelet packet basis function is used; energy features of each node are extracted:
[0027]
[0028] And calculate the energy entropy:
[0029]
[0030] The energy characteristics, energy entropy, and time-domain statistics of each layer are combined to form a multidimensional feature vector. .
[0031] Preferably, the feature fusion in step S3 employs a multi-head attention mechanism, assuming the extracted multi-source feature matrix is... Attention weights are calculated as follows:
[0032]
[0033] in The learnable parameter matrix; the features after multi-head fusion are:
[0034]
[0035]
[0036] Where H represents the number of attention heads. This is for outputting the projection matrix.
[0037] Preferably, the improved fault prediction model in step S4 is an LSTM network with attention gating, and its gating calculation formula is as follows:
[0038]
[0039] Wherein is the attention gating coefficient. The attention output from the previous moment is updated; the cell state and hidden state are updated as follows:
[0040]
[0041] The failure probability output is: .
[0042] Preferably, the fault prediction model in step S4 further includes an XGBoost ensemble classifier, which is stacked and fused with the LSTM model; the objective function of XGBoost is:
[0043]
[0044] Among the regularization terms T is the number of leaf nodes, and w is the leaf weight vector; the second-order Taylor expansion of the objective function in the t-th iteration is:
[0045]
[0046] in The optimal leaf weight is:
[0047]
[0048] in The final fusion prediction is:
[0049]
[0050] in , It is obtained by optimizing the AUC metric on the validation set.
[0051] Preferably, the graded early warning in step S5 adopts a dynamic threshold method based on distribution drift sensing, assuming the fault probability prediction sequence is... The dynamic threshold is calculated as follows:
[0052]
[0053] in and Here, represents the mean and standard deviation of the fault probability sequence, respectively, and the distribution drift metric is ;
[0054]
[0055] Sum is an adjustable parameter; when P> An early warning is triggered at a certain time, and the warning level is determined by the ratio η = P / It is determined that η∈(1,1.5] is a yellow alert, η∈(1.5,2.0] is an orange alert, and η>2.0 is a red alert.
[0056] Preferably, the remaining useful life (RUL) prediction in step S4 adopts a degradation model based on the Wiener process, and the degradation amount... obey:
[0057]
[0058] in For standard Brownian motion, The drift coefficient, Let be the diffusion coefficient; the probability distribution of RUL is:
[0059]
[0060] in Let Φ(⋅) be the failure threshold, and Φ(⋅) be the cumulative distribution function of the standard normal distribution; the expected value of the RUL prediction is:
[0061]
[0062] when Maintenance recommendations will be triggered at any time.
[0063] Preferably, it also includes feature importance analysis based on SHAP values and an adaptive acquisition strategy for the prediction model. ,feature The SHAP value is:
[0064]
[0065] The importance of features is ranked as follows:
[0066]
[0067] Based on the classification of features: when > Increase the frequency of this type of data collection to f high ,when Maintain standard frequency f std ,when Reduce the sampling frequency to f low This enables dynamic optimization and allocation of collected resources.
[0068] An automated instrument fault prediction system based on big data analytics, used to implement the method described above, includes:
[0069] The data acquisition layer is used to acquire multi-source operating data of automated instruments in real time via the OPC UA protocol;
[0070] The data processing layer is used to perform missing value imputation, outlier detection, wavelet packet feature extraction, and attention feature fusion.
[0071] The predictive analytics layer includes a Stacking ensemble prediction engine that integrates attention-gated LSTM and XGBoost models, as well as a RUL computation module based on the Wiener process.
[0072] The early warning decision layer is used to generate tiered early warning information based on dynamic thresholds, and adaptively adjust the data collection strategy based on SHAP feature importance analysis to push the early warning information to user terminals.
[0073] Compared with the prior art, the present invention has the following beneficial effects:
[0074] 1. This invention extracts the time-frequency joint energy features and energy entropy of the signal through wavelet packet decomposition, and introduces a multi-head attention mechanism to adaptively weight and fuse multi-source heterogeneous features. Compared with traditional single feature extraction methods, the fused feature vector can more comprehensively characterize the health status of the instrument, thereby improving the accuracy of fault identification.
[0075] 2. This invention introduces an attention gating coefficient α in LSTM. t This allows the model to adaptively focus on the most critical time step for fault determination; at the same time, it captures the high-order interaction relationship between features through the second-order Taylor expansion objective function of XGBoost. The two models are fused through the Stacking strategy, taking into account both time series modeling ability and feature combination ability, and the predicted AUC value is improved compared with the single model.
[0076] 3. The dynamic threshold formula proposed in this invention can sense changes in data distribution in real time and adaptively adjust the warning boundary. Compared with the fixed threshold method, the false alarm rate is reduced by about 30%-40% while maintaining a high fault detection rate. The introduction of the Wiener degradation model not only outputs the fault probability, but also gives the probability distribution and expected value of the remaining service life, providing a quantitative basis for maintenance plan formulation and avoiding over-maintenance or under-maintenance.
[0077] 4. This invention quantifies the contribution of each feature to the prediction result by using SHAP values and dynamically adjusts the collection frequency of different data sources. While ensuring prediction accuracy, it can reduce data storage and transmission overhead by about 25%-35%, making it particularly suitable for industrial field scenarios where edge computing resources are limited. Attached Figure Description
[0078] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0079] Figure 1 This is a flowchart of the method of the present invention;
[0080] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation
[0081] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0082] This invention provides a method and system for predicting faults in automated equipment based on big data analysis. This embodiment uses a pressure transmitter in a chemical plant reactor as the prediction target to illustrate the specific implementation process of this invention.
[0083] Example 1: Fault Prediction Method for Automated Instruments
[0084] Step S1: Real-time acquisition of multi-source operating data from automated instruments
[0085] This embodiment first constructs a data acquisition layer. Using the OPC UA protocol, it acquires operational data from the pressure transmitter over a continuous 30-day period at a sampling frequency of 1Hz in real time. The acquired multi-source operational data specifically includes three categories:
[0086] Process variable data: pressure value, temperature value, flow rate value.
[0087] Environmental variable data: ambient temperature, ambient humidity, vibration amplitude.
[0088] Instrument status data: power supply voltage, internal temperature, cumulative running time, and self-test status code.
[0089] The collected raw data is formed into a time series matrix, which serves as input for subsequent processing. The data acquisition layer transmits the collected raw data to the data processing layer via a message queue.
[0090] Step S2: Data Preprocessing
[0091] After receiving the raw data, the data processing layer performs the following preprocessing operations:
[0092] 1. Missing value imputation: For data with fewer than 5 consecutive missing time points, linear interpolation is used for imputation.
[0093] 2. Outlier Detection and Imputation: This embodiment employs an improved Z-score algorithm based on a sliding window. The specific implementation is as follows:
[0094] Set the sliding window length W=100, and the threshold Z th = 3.5, adaptive coefficient α = 0.5, and zero-prevention constant ε = 1e-8. For the data at time point i... i Calculate the local mean within the window. and local standard deviation Calculate the coefficient of variation within the window. Therefore, the adaptive coefficient k = 1 + 0.5 * Calculate the improved Z-score: .like If the value is greater than 3.5, then it is determined that... i This is an outlier.
[0095] The search returns 10 valid data points before and after each anomaly, forming a set N(i). A weighted interpolation formula is used to fill in the gaps, with an attenuation coefficient τ set to 5. Valid values closer to the current anomaly have higher weights. For example, for an anomaly in pressure at time i, a weighted replacement is performed using nearby normal pressure values.
[0096] 3. Data normalization: The Min-Max normalization method is used for all completed data to map the data range to the interval [0, 1] to eliminate the influence of different units.
[0097] Step S3: Multidimensional Feature Extraction and Fusion
[0098] 1. Wavelet packet decomposition feature extraction: This embodiment extracts features from the preprocessed pressure signal. A 3-level (J=3) wavelet packet decomposition was performed using the 'db4' wavelet basis function. The wavelet packet coefficients of each terminal node were obtained after decomposition. Extract the energy features of each node. and energy entropy Simultaneously, the time-domain statistics of the original signal are calculated: mean μ, standard deviation σ, skewness (Skew), and kurtosis (Kurt). These features are then combined into an initial multidimensional feature vector. The dimensions are (8+8+4)=20.
[0099] 2. Multi-head attention feature fusion: The extracted multi-source feature matrix is fused together. The input is fed into the multi-head attention mechanism layer, where m is the time step (32) and d is the feature dimension (20). In this embodiment, the number of attention heads H=4, and the dimension d of each head is... k = d / H = 5. Learnable parameter matrix W Q W K W V ∈ R {20×20} Output projection matrix W O ∈ R {20×20} The weighted features of each head are obtained by calculating Attention(Q,K,V), and finally the results of the four heads are concatenated and passed through W. O Projection yields the fused composite feature vector F. fused ∈ R{32×20} This process can adaptively learn the correlations between different time steps and different feature dimensions, highlighting the feature combinations that contribute the most to fault prediction.
[0100] Step S4: Failure Probability and Remaining Useful Life Prediction
[0101] The comprehensive feature vector F obtained in step S3 fused Input into the predictive analytics layer.
[0102] 1. LSTM Model Predicting Fault Probability: This embodiment constructs a two-layer LSTM network with 128 hidden units per layer, incorporating attention gating. Attention gating coefficients. The calculation method is as follows ,in This is the output of the attention gating at the previous time step. This gating coefficient directly affects the forgetting gate. In the calculation, the model can dynamically adjust the degree of forgetting of historical state information. Finally, the failure probability is output through a fully connected layer and a sigmoid activation function. .
[0103] 2. XGBoost model predicts fault probability: The flattened F_fused data is then input into the XGBoost classifier. The XGBoost model uses 100 trees with a maximum depth of 6 and a learning rate of 0.1. Its objective function includes a loss function l(y) that measures the prediction error. i , ŷ i The weights and regularization term Ω(f) are efficiently solved using a second-order Taylor expansion. During model training, the optimal weights are calculated for each leaf node. Where λ=1 is the L2 regularization coefficient. Output fault probability. .
[0104] 3. Stacking Model Fusion: LSTM and XGBoost are used as base learners, and logistic regression is used as the meta-learner for stacking fusion. Specifically, on an independent validation set, LSTM and XGBoost are used as base learners, and logistic regression is used as the meta-learner for stacking fusion. and As input features, a logistic regression model is trained to optimize the AUC metric, yielding the optimal fusion weights w1 and w2. The final fusion prediction result is:
[0105] .
[0106] In this embodiment, the validation set optimization yielded w1 = 0.65 and w2 = 0.35, indicating that the LSTM model contributes more to the modeling of temporal dependencies.
[0107] 4. RUL Prediction: A degradation model based on the Wiener process is adopted. The instrument drift is selected as the degradation factor. their obedience The drift coefficient is updated in real time using the maximum likelihood estimation method based on historical degradation data. and diffusion coefficient Set a failure threshold. This represents the critical value at which the instrument's accuracy exceeds tolerance. The degradation amount is given at the current time t. The expected value of the remaining useful life (RUL) is then... .when At that time, a maintenance suggestion will be triggered.
[0108] Step S5: Generate tiered early warning information
[0109] The early warning decision-making layer employs a dynamic threshold method based on distribution drift sensing. Let the fault probability prediction sequence, i.e., the prediction results for the most recent N=1000 time points, be... The dynamic threshold is calculated as follows:
[0110]
[0111] in, and These are the mean and standard deviation of the current failure probability sequence, respectively; β and γ are adjustable parameters, in this embodiment β=2 and γ=0.5; Drift is the distribution drift metric, defined as the degree of deviation between the mean failure probability of the current operating condition and the mean of the baseline operating condition.
[0112] The warning level is based on the ratio η = P current / Confirm and trigger the corresponding action:
[0113] Yellow warning (mild): Triggered when 1 < η ≤ 1.5. The system highlights the instrument on the monitoring interface and recommends scheduling an inspection within 72 hours.
[0114] Orange Moderate Warning: Triggered when 1.5 < η ≤ 2.0. The system sends an immediate message to the on-site engineer, recommending maintenance within 24 hours.
[0115] Red Severe Warning: Triggered when η > 2.0. The system issues an audible and visual alarm and automatically generates a maintenance work order, requiring immediate shutdown for repair.
[0116] When the model predicts a failure probability of 0.85 at a certain moment, and the current dynamic threshold... If the calculated value is 0.52, then η = 0.85 / 0.52 ≈ 1.63, and the system determines it to be an orange alert.
[0117] Furthermore, during operation, this system periodically analyzes feature importance based on SHAP values. For the prediction model... Calculate each feature SHAP value The average absolute SHAP value is then used as the feature importance score. .
[0118] Set an importance threshold = 0.05, = 0.01. According to Dynamically adjust the data acquisition frequency for the next cycle:
[0119] like >0.05, sampling frequency increased to f high = 2Hz.
[0120] If 0.01 0.05, maintaining the standard frequency f std = 1Hz.
[0121] like ≤ 0.01, sampling frequency reduced to f low = 0.1Hz.
[0122] By employing this strategy, this embodiment reduces data storage by approximately 30% while maintaining prediction accuracy, effectively saving resources for edge computing nodes.
[0123] Example 2:
[0124] like Figure 2 The system architecture diagram shown illustrates that the automated instrument fault prediction system based on big data analysis described in this invention specifically includes the following four layers for implementing the above method:
[0125] The data acquisition layer, deployed in the edge gateway of the industrial site, communicates with field automation instruments via the OPC UA protocol for non-intrusive real-time data acquisition. It also supports industrial protocols such as Modbus TCP / RTU. The acquired data undergoes preliminary formatting and timestamp alignment locally before being uploaded to the data processing layer.
[0126] The data processing layer, deployed on edge servers or in the cloud, receives raw data from the data acquisition layer and performs the following functions: a preprocessing module, which performs missing value imputation, improved Z-score outlier detection and weighted interpolation imputation, and data normalization; a feature engineering module, which performs wavelet packet decomposition, extraction of time-frequency domain energy features and energy entropy, and calculation of time-domain statistics; and a feature fusion module, which performs feature fusion using a multi-head attention mechanism and outputs a comprehensive feature vector.
[0127] The predictive analytics layer, deployed on a high-performance server or cloud platform, contains an integrated prediction engine consisting of the following core modules: an attention-gated LSTM model, which implements an improved LSTM network to capture temporal dependencies and output failure probabilities; an XGBoost model, which implements a gradient boosting tree model to capture high-order interactions between features; a Stacking fusion module, which implements a model fusion strategy to combine the outputs of LSTM and XGBoost to obtain the final failure probability; and a Wiener degradation module, which implements a remaining lifetime prediction model and outputs the expected RUL value and its probability distribution.
[0128] The early warning decision layer, deployed on the application server, is responsible for human-computer interaction and decision output. This layer includes: a dynamic threshold early warning module, which implements a dynamic threshold calculation method based on distribution drift perception to generate yellow, orange, and red three-level early warning information; an adaptive data acquisition strategy module, which implements feature importance analysis based on SHAP values, dynamically adjusts the data acquisition frequency of the data acquisition layer, and distributes the strategy to the data acquisition layer; and a user terminal push module, which pushes tiered early warning information, RUL prediction results, and maintenance suggestions to operations and maintenance personnel in real time through web interfaces, mobile apps, or SMS.
[0129] Taking the pressure transmitter in the aforementioned chemical plant as an example, the system operates according to the following process on the 20th day of operation:
[0130] The data acquisition layer collects data such as pressure, temperature, and voltage at a frequency of 1Hz.
[0131] The data processing layer detected a sudden change in pressure to 10.5 MPa at a certain moment, while the mean within the window was 9.0 MPa, resulting in an improved Z-score of 4.2, which was identified as an outlier. The system corrected it to 9.05 MPa using weighted interpolation of adjacent effective values and then normalized it.
[0132] The feature extraction module performs wavelet packet decomposition on the pressure signal and discovers the high-frequency energy feature E. {3,7} It is significantly higher than the historical baseline; the multi-head attention mechanism assigns a high weight of 0.8 to this feature.
[0133] The LSTM model outputs a fault probability of 0.88, the XGBoost model outputs 0.75, and the final fault probability after fusion is 0.65*0.88 + 0.35*0.75 = 0.8345.
[0134] The RUL prediction module calculates E[RUL] to be approximately 35 hours based on the current cumulative drift amount of 0.12mA and the failure threshold of 0.5mA.
[0135] The early warning decision-making layer calculates the current dynamic threshold. = 0.4, η = 0.8345 / 0.4 = 2.086 > 2.0, immediately triggering a red alert and generating an alarm message: "The pressure transmitter is about to experience an accuracy error, with an estimated remaining service life of 35 hours. Immediate shutdown and maintenance are recommended." The on-duty engineer is notified via SMS and the monitoring screen. Simultaneously, SHAP analysis shows that the importance of "pressure value" and "supply voltage" exceeds the threshold, and the system automatically increases the acquisition frequency of these two data points to 2Hz.
[0136] The above description of specific implementation methods clearly and completely demonstrates the automated equipment fault prediction method and system based on big data analysis proposed in this invention. This method effectively solves problems in existing technologies such as insufficient feature extraction, weak generalization ability of prediction models, rigid warning thresholds, and unreasonable resource utilization, significantly improving the intelligence level and practical value of predictive maintenance for automated instruments.
[0137] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.
Claims
1. A method for predicting faults in automated equipment based on big data analysis, characterized in that: Includes the following steps: S1: Real-time acquisition of multi-source operating data from automated instruments, including process variable data, environmental variable data, and instrument status data; S2: Preprocess the multi-source operating data, including missing value imputation, outlier detection, and data normalization; S3: Based on the preprocessed data, perform multi-dimensional feature extraction and fusion to obtain a comprehensive feature vector; S4: Input the comprehensive feature vector into the improved fault prediction model and output the fault probability and remaining service life prediction values; S5: Generate graded early warning information based on the failure probability and remaining service life prediction.
2. The method for predicting faults in automated equipment based on big data analysis according to claim 1, characterized in that: The outlier detection in step S2 employs an improved Z-score algorithm based on a sliding window, specifically including: Set the sliding window length to W, for the data at the i-th time point i Calculate local statistics within the window: The improved Z-score is as follows: Where k is the adaptive coefficient. The coefficient of variation within the window. To prevent division by zero constant; when >Z th If a value is identified as an outlier, it is filled using weighted interpolation of adjacent valid values. The filling formula is as follows: in Let i be the set of valid samples surrounding anomaly point i. This is the attenuation coefficient.
3. The method for predicting faults in automated equipment based on big data analysis according to claim 1, characterized in that: The multidimensional feature extraction in step S3 employs wavelet packet decomposition to extract the signal. Perform J-level wavelet packet decomposition: in Let be the wavelet packet coefficients of the k-th node in the j-th layer. The wavelet packet basis function is used; energy features of each node are extracted: And calculate the energy entropy: The energy characteristics, energy entropy, and time-domain statistics of each layer are combined to form a multidimensional feature vector. .
4. The method for predicting faults in automated equipment based on big data analysis according to claim 3, characterized in that: The feature fusion described in step S3 employs a multi-head attention mechanism. Let the extracted multi-source feature matrix be... Attention weights are calculated as follows: in The learnable parameter matrix; the features after multi-head fusion are: Where H represents the number of attention heads. This is for outputting the projection matrix.
5. The method for predicting faults in automated equipment based on big data analysis according to claim 1, characterized in that: The improved fault prediction model described in step S4 is an LSTM network with attention gating, and its gating calculation formula is as follows: Wherein is the attention gating coefficient. The attention output from the previous moment is updated; the cell state and hidden state are updated as follows: The failure probability output is: .
6. The method for predicting faults in automated equipment based on big data analysis according to claim 1, characterized in that: The fault prediction model in step S4 also includes an XGBoost ensemble classifier, which is stacked and fused with the LSTM model; the objective function of XGBoost is: Among the regularization terms T is the number of leaf nodes, and w is the leaf weight vector; the second-order Taylor expansion of the objective function in the t-th iteration is: in The optimal leaf weight is: in The final fusion prediction is: in , It is obtained by optimizing the AUC metric on the validation set.
7. The method for predicting faults in automated equipment based on big data analysis according to claim 1, characterized in that: The graded early warning method described in step S5 employs a dynamic threshold method based on distribution drift sensing, assuming the fault probability prediction sequence is... The dynamic threshold is calculated as follows: in and Here, represents the mean and standard deviation of the fault probability sequence, respectively, and the distribution drift metric is ; Sum is an adjustable parameter; when P> An early warning is triggered at a certain time, and the warning level is determined by the ratio η=P / It is determined that η∈(1,1.5] is a yellow alert, η∈(1.5,2.0] is an orange alert, and η>2.0 is a red alert.
8. The method for predicting faults in automated equipment based on big data analysis according to claim 1, characterized in that: The remaining useful life (RUL) prediction in step S4 uses a degradation model based on the Wiener process, and the degradation amount... obey: in For standard Brownian motion, The drift coefficient, Let be the diffusion coefficient; the probability distribution of RUL is: in Let Φ(⋅) be the failure threshold, and Φ(⋅) be the cumulative distribution function of the standard normal distribution; the expected value of the RUL prediction is: when Maintenance recommendations will be triggered at any time.
9. The method for predicting faults in automated equipment based on big data analysis according to claim 1, characterized in that: It also includes feature importance analysis based on SHAP values and adaptive acquisition strategies for prediction models. ,feature The SHAP value is: The importance of features is ranked as follows: Based on the classification of features: when > Increase the frequency of this type of data collection to f high ,when Maintain standard frequency f std ,when Reduce the sampling frequency to f low This enables dynamic optimization and allocation of collected resources.
10. An automated instrument fault prediction system based on big data analysis, used to implement the method described in any one of claims 1-9, characterized in that: include: The data acquisition layer is used to acquire multi-source operating data of automated instruments in real time via the OPC UA protocol; The data processing layer is used to perform missing value imputation, outlier detection, wavelet packet feature extraction, and attention feature fusion. The predictive analytics layer includes a Stacking ensemble prediction engine that integrates attention-gated LSTM and XGBoost models, as well as a RUL computation module based on the Wiener process. The early warning decision layer is used to generate tiered early warning information based on dynamic thresholds, and adaptively adjust the data collection strategy based on SHAP feature importance analysis to push the early warning information to user terminals.