Failure assessment method and device for intelligent control system of coal wharf

Abstract

The invention provides a coal wharf intelligent control system failure assessment method and device, and the method comprises the steps: setting a basic window length according to an equipment type, calculating a window length offset through combining historical fault data, and obtaining an optimized window length; extracting time sequence characteristics such as an autocorrelation coefficient and a partial autocorrelation coefficient from the standardized time sequence data corresponding to the optimized window length; and determining a health index and a residual service life prediction value according to the time sequence characteristics by utilizing a long-short-term memory network and a prediction model of a Transform structure. According to the method, the window length is dynamically calculated and optimized, the key time sequence characteristics are extracted, and the specific prediction model is adopted, so that the real-time and accurate evaluation on the failure of the intelligent control system of the coal wharf is realized, and the defects of the existing method in the aspects of processing the dynamic evolution and prediction splitting of multi-source heterogeneous data are effectively overcome.

Description

Technical Field

[0001] This application belongs to the field of control failure assessment, and in particular relates to a failure assessment method and device for an intelligent control system of a coal terminal. Background Technology

[0002] Traditional failure assessment of intelligent control systems at coal terminals relies on periodic maintenance and human experience to monitor and maintain system operation, ensuring loading and unloading efficiency and equipment safety. This approach depends on fixed maintenance cycles and operator subjective judgment, aiming to achieve basic condition monitoring and fault prevention.

[0003] Specifically, this includes regular maintenance plans and simple data analysis processes. Operational data is processed through manual inspections or static models. For example, using only a single data source (such as motor current) for fault diagnosis can achieve basic condition assessment results. Some data-driven methods use fixed-parameter models to analyze electrical signals, achieving limited fault identification, but they cannot integrate the dynamic evolution of multi-source heterogeneous time-series data (such as electrical signals, mechanical vibrations, and environmental parameters).

[0004] The static or fragmented processing methods rely on fixed windows or single models, which cannot adapt to the dynamic changes and coupling characteristics of multi-source heterogeneous time-series data. This leads to difficulties in identifying early fault symptoms, inaccurate prediction of remaining lifetime, biased evaluation results, and limited decision support capabilities. Summary of the Invention

[0005] The purpose of this application is to overcome the deficiencies in the prior art and provide a method and apparatus for failure assessment of an intelligent control system for coal terminals.

[0006] This application provides a failure assessment method for an intelligent control system at a coal terminal, including:

[0007] The basic window length is set according to the device type, and the window length offset is calculated based on historical fault data. The optimized window length is then calculated based on the basic window length and the window length offset.

[0008] Temporal features are extracted from the standardized time series data corresponding to the optimized window length, and the temporal features include autocorrelation coefficients and partial autocorrelation coefficients.

[0009] Based on the aforementioned temporal characteristics, a prediction model using a long short-term memory network and a Transformer structure is used to determine the predicted values ​​of the health index and remaining lifespan.

[0010] Optionally, the window length offset is calculated based on historical fault data, including:

[0011] The range of values ​​for the proportional coefficient 'a' is determined based on the equipment type classification.

[0012] The window length offset is calculated based on the scaling factor a and the mean fault interval time (MTBF) in the historical fault data.

[0013] Optionally, extracting time-series features from the standardized time-series data corresponding to the optimized window length includes:

[0014] Calculate the autocorrelation coefficient and the partial autocorrelation coefficient;

[0015] Generate a feature vector set, which includes the autocorrelation coefficient, the partial autocorrelation coefficient, the device ID, and the timestamp.

[0016] Optionally, predictive models using long short-term memory networks and Transformer architectures include:

[0017] The encoder uses a bidirectional LSTM layer with multiple hidden units;

[0018] The bidirectional LSTM layer is used to capture long-term dependencies in the input sequence.

[0019] Optionally, the method further includes:

[0020] Use a short-term diagnostic model to determine the failure probability distribution;

[0021] A risk assessment level is generated based on the health index and the maximum failure probability in the failure probability distribution;

[0022] Based on the risk assessment level, maintenance decision recommendations are generated;

[0023] The short-term diagnostic model is a neural network model based on an attention mechanism.

[0024] This application also provides a failure assessment device for an intelligent control system at a coal terminal, comprising:

[0025] The window module sets a basic window length based on the device type, calculates the window length offset based on historical fault data, and calculates an optimized window length based on the basic window length and the window length offset.

[0026] The time series module extracts time series features from the standardized time series data corresponding to the optimized window length. The time series features include autocorrelation coefficients and partial autocorrelation coefficients.

[0027] The prediction module determines the predicted values ​​of health index and remaining lifespan using a prediction model with a long short-term memory network and a Transformer structure, based on the time-series characteristics.

[0028] Optionally, the window module calculates the window length offset based on historical fault data, including:

[0029] The range of values ​​for the proportional coefficient 'a' is determined based on the equipment type classification.

[0030] The window length offset is calculated based on the scaling factor a and the mean fault interval time (MTBF) in the historical fault data.

[0031] Optionally, the time series module extracts time series features from the standardized time series data corresponding to the optimized window length, including:

[0032] Calculate the autocorrelation coefficient and the partial autocorrelation coefficient;

[0033] Generate a feature vector set, which includes the autocorrelation coefficient, the partial autocorrelation coefficient, the device ID, and the timestamp.

[0034] Optionally, the prediction module uses a prediction model with a long short-term memory network and a Transformer structure, including:

[0035] The encoder uses a bidirectional LSTM layer with multiple hidden units;

[0036] The bidirectional LSTM layer is used to capture long-term dependencies in the input sequence.

[0037] Optionally, the device further includes:

[0038] The decision-making module uses a short-term diagnostic model to determine the failure probability distribution; generates a risk assessment level based on the health index and the maximum failure probability in the failure probability distribution; and generates maintenance decision recommendations based on the risk assessment level; wherein the short-term diagnostic model is a neural network model based on an attention mechanism.

[0039] The beneficial effects of this application are:

[0040] This application provides a failure assessment method for an intelligent control system at a coal terminal, comprising: setting a basic window length based on equipment type, calculating a window length offset based on historical fault data, and calculating an optimized window length based on the basic window length and the window length offset; extracting time-series features from standardized time-series data corresponding to the optimized window length, the time-series features including autocorrelation coefficients and partial autocorrelation coefficients; and determining a health index and predicted remaining useful life using a prediction model with a long short-term memory network and a Transformer structure based on the time-series features. This application achieves real-time and accurate failure assessment of intelligent control systems at coal terminals by setting a basic window length based on equipment type and dynamically calculating an optimized window length based on historical fault data, extracting time-series features such as autocorrelation coefficients and partial autocorrelation coefficients from standardized time-series data, and using a prediction model with a long short-term memory network and a Transformer structure to determine the health index and predicted remaining useful life. This overcomes the shortcomings of existing methods in handling the dynamic evolution and fragmented prediction of multi-source heterogeneous data. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the failure assessment process for the intelligent control system of a coal terminal.

[0042] Figure 2 This is a schematic diagram of the failure assessment device for the intelligent control system of a coal terminal. Detailed Implementation

[0043] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is to be understood that various forms of implementation of the present disclosure are intended and should not be limited to the embodiments set forth herein. Rather, the embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0044] Please refer to Figure 1 As shown, this application provides a machine learning-based failure assessment method for intelligent control systems in coal terminals, applied in the field of intelligent control of coal terminals, to solve the failure assessment problem of intelligent control systems in coal terminals. The method includes:

[0045] S101. Set the basic window length according to the device type, calculate the window length offset based on historical fault data, and calculate the optimized window length based on the basic window length and the window length offset.

[0046] The data acquisition terminal collects real-time operational status data and environmental data from various devices in the intelligent control system of the coal terminal, forming a multi-source heterogeneous time-series dataset. Operational status data includes motor current, voltage, speed, torque, and valve opening signals from the intelligent control system. Environmental data includes wind speed, humidity, and dust concentration.

[0047] The data acquisition terminal is deployed near the intelligent control system of the coal terminal and has multi-protocol access capabilities, including Modbus, OPC UA, and Profibus. The intelligent control system of the coal terminal is connected via industrial Ethernet to read motor current, voltage, speed, torque, and valve opening signals in real time. Environmental data sources include wind speed sensors, humidity sensors, and dust concentration monitors, which are connected to the data acquisition terminal via RS485 / Modbus or 4-20mA analog signals.

[0048] The data acquisition method adopts high-frequency polling and time-triggered real-time acquisition. Independent acquisition threads are set up for different data sources, and data is collected asynchronously and in multiple threads. Each data acquisition is accompanied by a high-precision timestamp to provide a basis for subsequent data alignment.

[0049] The aforementioned high-frequency polling refers to the data acquisition terminal actively requesting data from sensors or control systems at a short period (e.g., seconds or milliseconds) that is much faster than the normal rate of change in device status. Time-triggered methods refer to initiating data acquisition at a specific, predefined time point or when a certain time condition is met.

[0050] Independent acquisition threads refer to creating dedicated, non-interfering execution paths for data acquisition for each data type or data source (such as current sensors or vibration sensors). Asynchronous multi-threaded acquisition means that these threads run independently without waiting for other threads to complete, and the acquired data is placed in a shared queue or buffer for unified management.

[0051] Motor current is measured in A, voltage in V, speed in RPM, torque in Nm, valve opening in %, wind speed in m / s, humidity in %RH, and dust concentration in mg / m³. Each data entry in the data structure includes device ID, signal type, timestamp, and value.

[0052] The data processing terminal cleans, aligns, and dynamically normalizes the collected multi-source heterogeneous time-series datasets using a sliding time window, forming a standardized time-series dataset. Cleaning operations include missing value handling, outlier handling, and duplicate data removal. Alignment operations include timestamp synchronization and sampling frequency unification. Outlier detection and handling employ the 3σ principle to identify outliers, which are then removed before interpolation. Missing value handling uses the mean within the sliding window for filling. Duplicate data removal is based on timestamp deduplication, retaining the latest valid values.

[0053] Timestamp synchronization is based on the clock of the intelligent control system of the coal terminal, aligning all sensor data. Linear interpolation is used to unify data with different sampling frequencies onto the same time axis. Sampling frequency unification involves downsampling high-frequency data and upsampling low-frequency data to a unified frequency using resampling methods such as pandas.resample() to achieve frequency alignment.

[0054] Dynamic normalization is based on the formula:

[0055]

[0056] in, Here, X represents the normalized data, and X represents the original data collected so far. and These are the minimum and maximum values ​​of the feature data within a sliding window of a preset time length prior to the current moment.

[0057] The length of the sliding time window is dynamically adjusted based on the equipment type and historical fault data. A base window length is set according to the equipment type; for critical, high-speed, or fault-prone equipment, such as large conveyor motors and crusher spindle motors, whose status changes rapidly and faults have a significant impact, requiring a quick response, a shorter window length is set. For example, 1 hour; ordinary or low-speed or auxiliary equipment, such as ordinary pumps, lighting control systems, with slow status changes and longer settings. For example, 12 hours or 24 hours; the frequency of environmental parameter changes is between these two, and can be set to 4 hours or 8 hours.

[0058] Calculate the historical fault data offset ΔT, analyze the historical fault data of this device or similar devices, and calculate its mean time between failures (MTBF):

[0059]

[0060] Where 'a' is a proportionality coefficient, an empirical value given based on the criticality of the equipment and the stability of historical data. For critical, fault-prone equipment, the value of 'a' ranges from [-0.02, -0.01], for critical, stable equipment, the value of 'a' ranges from [-0.001, 0], for ordinary equipment, the value of 'a' ranges from [0, 0.005], and for environmental monitoring sensors, 'a' is 0. If historical data shows that the equipment fails frequently and has a short MTBF, then ΔT is negative, and the window length should be shortened to more sensitively capture its rapidly changing fault characteristics. If historical data shows that the equipment operates stably and has a long MTBF, then ΔT is positive, and the window length should be extended to better fit its long-term stable operating baseline and smooth out short-term fluctuations.

[0061] Calculate the final window length According to the formula:

[0062]

[0063] To prevent windows from becoming too long or too short, which could lead to a loss of dynamism, excessive sensitivity, or excessive noise, upper and lower limits are set for window length. and , minute, Heaven; if ,but ,if ,but .

[0064] S102. Extract time series features from the standardized time series data corresponding to the optimized window length. The time series features include autocorrelation coefficient and partial autocorrelation coefficient.

[0065] The feature extraction terminal extracts time-domain features, frequency-domain features, and time-series features from standardized time-series data, generating a feature vector set for model input. Time-domain features include mean, variance, peak value, and kurtosis. Frequency-domain features include spectral peaks and major frequency components obtained through Fast Fourier Transform.

[0066] The maxima factor is the ratio of the signal's peak value to its root mean square value, used to measure the intensity of the impulse component in the signal. Kurtosis is a statistic describing the difference between the shape of the signal's probability distribution and the normal distribution, reflecting the number of extreme events in the signal.

[0067] Time series characteristics include autocorrelation coefficient and partial autocorrelation coefficient. The autocorrelation coefficient measures the linear correlation between observations of a time series at different time points. It describes the similarity between data at the current moment and data at a past moment. The partial autocorrelation coefficient, on the other hand, measures the direct correlation between data at the current moment and data at a specific past moment, excluding the influence of intermediate time points.

[0068] The generated structured feature vector set contains device ID, timestamp, and feature vector, with each feature vector corresponding to the state representation of a device within a certain time window.

[0069] S103. Based on the temporal characteristics, use a prediction model with a long short-term memory network and a Transformer structure to determine the predicted values ​​of the health index and remaining lifespan.

[0070] The feature vector set containing time-domain, frequency-domain, and time-series features (autocorrelation coefficient, partial autocorrelation coefficient) is input into the pre-trained short-term diagnostic model and long-term prediction model, which output the probability distribution of the current fault type, the health index, and the predicted value of the remaining service life, respectively. The short-term diagnostic model is a deep convolutional neural network model based on an attention mechanism, and the long-term prediction model is a time-series prediction model combining LSTM and Transformer structures.

[0071] Among them, the autocorrelation coefficient helps the model perceive the periodic degradation pattern of equipment operating status, while the partial autocorrelation coefficient can provide clues for judging the intrinsic correlation between faults. The model's encoder (LSTM) first encodes these sequential features to capture their long-term dependencies; then, the Transformer module uses a self-attention mechanism to dynamically weight the feature contributions at different times, paying particular attention to the key temporal dependency patterns revealed by autocorrelation and partial autocorrelation, thereby more accurately determining the health index and remaining service life prediction values.

[0072] The specific network structure of the short-term diagnostic model is as follows: the input layer receives a feature vector of dimension n, where n is the number of features; the one-dimensional convolutional layer group contains two convolutional layers: one-dimensional convolutional layer 1 has a kernel size of 3, 64 channels, and the activation function is ReLU; one-dimensional convolutional layer 2 has a kernel size of 5, 128 channels, and the activation function is ReLU; the attention mechanism layer adopts a multi-head self-attention mechanism with 8 heads, which is used to calculate the importance weights of features at different time steps in the sequence, focusing on key feature periods related to the fault; the global average pooling layer pools the feature sequences output by the convolutional and attention layers, converting them into fixed-length feature vectors; the fully connected output layer uses the Softmax activation function to output the probability distribution of all preset fault types.

[0073] The specific network structure of the long-term prediction model is that the input layer receives a sequence of feature vectors with the same structure as the short-term diagnostic model. The encoder first consists of a single bidirectional LSTM layer with 128 hidden units, used to encode the input sequence and capture long-term dependencies. The Transformer module receives the output sequence of the LSTM layer as input. This module contains an encoder layer with a multi-head self-attention mechanism at its core (4 heads), a feedforward neural network, and includes residual connections and layer normalization. The decoder or output layer uses the state of the last time step output by the Transformer encoder as input to a fully connected layer network for regression prediction. This network contains two branches:

[0074] The health index prediction branch consists of a fully connected layer with a dimension of 64 and a linear output layer, outputting a scalar health index value. The remaining useful life prediction branch consists of a fully connected layer with a dimension of 64 and a linear output layer, outputting a scalar RUL value.

[0075] The formula for calculating the health index is:

[0076]

[0077] in, Let i be the i-th feature in the feature vector set extracted from equipment operating status data and environmental data. and Let be the minimum and maximum allowable values ​​of the i-th feature under the device health baseline state, respectively. Let be the weight coefficients corresponding to each feature, and satisfy . The formula for calculating the predicted remaining useful life is:

[0078]

[0079] in, Here, |d(HI) / dt| is a predefined equipment failure threshold, and |d(HI) / dt| is the instantaneous rate of decline of the health index.

[0080] The decision mapping terminal, based on preset rules, integrates the outputs of short-term diagnostic models and long-term prediction models to generate a comprehensive failure risk assessment level and maintenance decision recommendations. The comprehensive failure risk assessment level is generated based on a health index using a formula.

[0081]

[0082] in, and The risk threshold of the health index and The risk threshold representing the probability of failure. for.

[0083] Maintenance decision recommendations are generated based on the failure risk assessment level, including "continue operation and regular monitoring" for low risk level; "planned maintenance" for medium risk level; and "immediate shutdown and maintenance" for high risk level.

[0084] After implementing maintenance decision recommendations, the actual results data are fed back to the evaluation model, and the model is optimized based on machine learning algorithms; the machine learning algorithm is an incremental learning algorithm.

[0085] After implementing maintenance decision recommendations, the system collects the following actual result data to form a feedback dataset: device ID, signal type, timestamp, maintenance operation type, post-maintenance device status data, operating status and environmental data defined in the data acquisition, actual fault type, health index change trend, and previous prediction results of the evaluation model; using incremental learning algorithms, such as Online Random Forest, Incremental SVM, or Online Deep Learning, the evaluation model is updated according to the following process:

[0086] The feedback data undergoes the same cleaning, alignment, and feature extraction processes as the data processing and feature extraction.

[0087] The samples are labeled with their true labels based on actual maintenance results, such as fault type and health index; the short-term diagnostic model is fine-tuned online using new samples;

[0088] The long-term prediction model is trained using incremental sequence training with new sequence data; the feature weights are dynamically adjusted based on the contribution of each feature in the feedback data. .

[0089] like Figure 2 As shown, this application also provides a failure assessment device for an intelligent control system at a coal terminal, comprising:

[0090] Window module 201 sets a basic window length according to the device type, calculates a window length offset based on historical fault data, and calculates an optimized window length based on the basic window length and the window length offset.

[0091] The time series module 202 extracts time series features from the standardized time series data corresponding to the optimized window length. The time series features include autocorrelation coefficients and partial autocorrelation coefficients.

[0092] The prediction module 203 determines the predicted values ​​of the health index and remaining lifespan using a prediction model with a long short-term memory network and a Transformer structure, based on the time-series characteristics.

[0093] Optionally, the window module calculates the window length offset based on historical fault data, including:

[0094] The range of values ​​for the proportional coefficient 'a' is determined based on the equipment type classification.

[0095] The window length offset is calculated based on the scaling factor a and the mean fault interval time (MTBF) in the historical fault data.

[0096] Optionally, the time series module extracts time series features from the standardized time series data corresponding to the optimized window length, including:

[0097] Calculate the autocorrelation coefficient and the partial autocorrelation coefficient;

[0098] Generate a feature vector set, which includes the autocorrelation coefficient, the partial autocorrelation coefficient, the device ID, and the timestamp.

[0099] Optionally, the prediction module uses a prediction model with a long short-term memory network and a Transformer structure, including:

[0100] The encoder uses a bidirectional LSTM layer with multiple hidden units;

[0101] The bidirectional LSTM layer is used to capture long-term dependencies in the input sequence.

[0102] Optionally, the device further includes:

[0103] The decision module 204 uses a short-term diagnostic model to determine the failure probability distribution; generates a risk assessment level based on the health index and the maximum failure probability in the failure probability distribution; and generates maintenance decision recommendations based on the risk assessment level; wherein the short-term diagnostic model is a neural network model based on an attention mechanism.

[0104] The above description of the embodiments is provided to enable those skilled in the art to understand and apply this application. Those skilled in the art will readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without inventive effort. Therefore, this application is not limited to the above embodiments, and any improvements and modifications made to this application based on the disclosure thereof should be within the scope of protection of this application.

Claims

1. A failure assessment method for an intelligent control system at a coal terminal, characterized in that, include: The basic window length is set according to the device type, and the window length offset is calculated based on historical fault data. The optimized window length is then calculated based on the basic window length and the window length offset. Temporal features are extracted from the standardized time series data corresponding to the optimized window length, and the temporal features include autocorrelation coefficients and partial autocorrelation coefficients. Based on the aforementioned temporal characteristics, a prediction model using a long short-term memory network and a Transformer structure is used to determine the predicted values ​​of the health index and remaining lifespan. The step of setting the basic window length according to the device type includes: classifying devices into critical devices, general devices, and environmental monitoring sensors, and setting the basic window length accordingly. The formula for calculating the window length offset is as follows: Where MTBF is the average fault interval time in historical fault data, and 'a' is the scaling factor; the optimization window length is: ; The health index is expressed by the following formula: ,in The weight coefficients of each feature and , These are feature values ​​extracted from operational status data and environmental data. and The minimum and maximum permissible values ​​for the feature under healthy baseline conditions; the predicted remaining useful life (RUL) is expressed by the formula... ,in This is the equipment failure threshold. This represents the instantaneous rate of decline in the health index.

2. The method according to claim 1, characterized in that, The window length offset is calculated based on historical fault data, including: The range of values ​​for the proportional coefficient 'a' is determined based on the equipment type classification. The window length offset is calculated based on the scaling factor a and the mean fault interval time (MTBF) in the historical fault data.

3. The method according to claim 1, characterized in that, Extracting time series features from the standardized time series data corresponding to the optimized window length includes: Calculate the autocorrelation coefficient and the partial autocorrelation coefficient; Generate a feature vector set, which includes the autocorrelation coefficient, the partial autocorrelation coefficient, the device ID, and the timestamp.

4. The method according to claim 1, characterized in that, Predictive models using Long Short-Term Memory networks and Transformer architectures include: The encoder uses a bidirectional LSTM layer with multiple hidden units; The bidirectional LSTM layer is used to capture long-term dependencies in the input sequence.

5. The method according to claim 1, characterized in that, The method further includes: Use a short-term diagnostic model to determine the failure probability distribution; A risk assessment level is generated based on the health index and the maximum failure probability in the failure probability distribution; Based on the risk assessment level, maintenance decision recommendations are generated; The short-term diagnostic model is a neural network model based on an attention mechanism.

6. A failure assessment device for an intelligent control system at a coal terminal, characterized in that, include: The window module sets a basic window length based on the device type, calculates the window length offset based on historical fault data, and calculates an optimized window length based on the basic window length and the window length offset. The time series module extracts time series features from the standardized time series data corresponding to the optimized window length. The time series features include autocorrelation coefficients and partial autocorrelation coefficients. The prediction module determines the predicted values ​​of health index and remaining lifespan using a prediction model with a long short-term memory network and a Transformer structure, based on the time-series characteristics. The step of setting the basic window length according to the device type includes: classifying devices into critical devices, general devices, and environmental monitoring sensors, and setting the basic window length accordingly. The formula for calculating the window length offset is as follows: Where MTBF is the average fault interval time in historical fault data, and 'a' is the scaling factor; the optimization window length is: ; The health index is expressed by the following formula: ,in The weight coefficients of each feature and , These are feature values ​​extracted from operational status data and environmental data. and The minimum and maximum permissible values ​​for the feature under healthy baseline conditions; the predicted remaining useful life (RUL) is expressed by the formula... ,in This is the equipment failure threshold. This represents the instantaneous rate of decline in the health index.

7. The apparatus according to claim 6, characterized in that, The window module calculates the window length offset based on historical fault data, including: The range of values ​​for the proportional coefficient 'a' is determined based on the equipment type classification. The window length offset is calculated based on the scaling factor a and the mean fault interval time (MTBF) in the historical fault data.

8. The apparatus according to claim 6, characterized in that, The timing module extracts timing features from the standardized timing data corresponding to the optimized window length, including: Calculate the autocorrelation coefficient and the partial autocorrelation coefficient; Generate a feature vector set, which includes the autocorrelation coefficient, the partial autocorrelation coefficient, the device ID, and the timestamp.

9. The apparatus according to claim 6, characterized in that, The prediction module uses a prediction model based on a long short-term memory network and a Transformer architecture, including: The encoder uses a bidirectional LSTM layer with multiple hidden units; The bidirectional LSTM layer is used to capture long-term dependencies in the input sequence.

10. The apparatus according to claim 6, characterized in that, The device further includes: The decision-making module uses a short-term diagnostic model to determine the failure probability distribution; generates a risk assessment level based on the health index and the maximum failure probability in the failure probability distribution; and generates maintenance decision recommendations based on the risk assessment level; wherein the short-term diagnostic model is a neural network model based on an attention mechanism.

Images