An atmospheric tower corrosion fault intelligent diagnosis method and system based on deep learning and mechanism constraint fusion

Abstract

The present application relates to a kind of atmospheric tower corrosion fault intelligent diagnosis method and system based on deep learning and mechanism constraint fusion.The method is by collecting the wall thickness, temperature, pressure, vibration and medium composition data of multiple spatial positions of atmospheric tower, constructs space-time data sequence, after pre-processing and multi-source fusion, utilizes convolutional neural network and long short-term memory network to extract corrosion evolution characteristics, corrosion type identification, corrosion rate prediction and residual service life evaluation are realized in combination with deep neural network and time series prediction model, and constraint correction is carried out based on mechanism formula, while realizing early warning evaluation and model incremental updating.The present application can improve the accuracy, reliability and intelligent level of atmospheric tower corrosion diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for petrochemical equipment, specifically to an intelligent diagnostic method and system for corrosion faults in atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints, belonging to the field of industrial equipment condition monitoring, fault diagnosis, and predictive maintenance technology. Background Technology

[0002] In the petrochemical production process, atmospheric distillation units are one of the core equipment in the refining system. Atmospheric towers operate in high-temperature, high-pressure, and multiphase flow environments for extended periods. The media typically contain corrosive components such as sulfides and naphthenic acids, which can easily corrode critical components such as the tower walls, trays, downcomers, and inlet / outlet ports. This can lead to various corrosion failure modes, including uniform corrosion, pitting, erosion corrosion, and stress corrosion cracking. In severe cases, this can result in equipment leaks, unplanned shutdowns, or even safety accidents.

[0003] In existing technologies, the monitoring and diagnosis of corrosion problems in atmospheric pressure towers typically employ a combination of manual inspections and periodic offline testing. For example, ultrasonic thickness measurements are performed on critical components periodically, or comprehensive inspections are conducted during unit shutdowns for maintenance to assess the degree of corrosion. While this approach can reflect the corrosion status of the equipment to some extent, it still has significant limitations.

[0004] First, existing monitoring methods mostly involve intermittent data acquisition, lacking continuity and failing to reflect the dynamic evolution characteristics of the corrosion process. This results in insufficient understanding of corrosion development trends and delayed fault detection. Second, traditional diagnostics rely heavily on the experience of engineers, which is highly subjective. It struggles to accurately identify corrosion types and evolution mechanisms caused by the coupling of multiple factors under complex operating conditions, affecting the reliability of diagnostic results.

[0005] Furthermore, existing technologies typically lack effective means of predicting corrosion rates and remaining service life, relying mostly on empirical estimations or simple models. This makes it difficult to accurately assess the future operating status of equipment, thus hindering scientific predictive maintenance decisions. Simultaneously, traditional methods lack the ability to fuse and analyze multi-source data (such as wall thickness, temperature, pressure, vibration, and media composition), failing to fully exploit the correlations between data points and resulting in low information utilization.

[0006] With the development of industrial IoT and artificial intelligence technologies, some studies have attempted to apply machine learning or deep learning methods to the field of equipment fault diagnosis. However, existing solutions are mostly focused on single data sources or single model applications, lacking effective integration with actual corrosion mechanisms, which easily leads to inconsistencies between model prediction results and physical laws. At the same time, most systems have failed to form a closed-loop mechanism for data acquisition, diagnostic analysis, prediction evaluation, and feedback optimization, resulting in insufficient model generalization ability and long-term adaptability.

[0007] Therefore, there is an urgent need for an intelligent diagnostic technology for corrosion faults in atmospheric pressure towers that can integrate multi-source spatiotemporal data, combine deep learning with corrosion mechanism constraints, and have continuous learning capabilities. This technology can enable real-time monitoring of corrosion status, accurate identification of fault types, reliable prediction of corrosion rate and remaining service life, and provide effective support for equipment safety operation and maintenance decisions. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide an intelligent diagnostic method and system for corrosion faults in atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints, to solve the problems of discontinuous corrosion monitoring data, reliance on manual experience for fault diagnosis, insufficient accuracy in predicting corrosion rate and remaining service life, and lack of closed-loop optimization capabilities in existing technologies. To achieve the above objective, this invention provides the following technical solution: In one embodiment of the present invention, an intelligent diagnostic method for corrosion faults in atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints is provided. Data on wall thickness, temperature, pressure, vibration, and media composition are collected at multiple spatial locations within the atmospheric pressure tower, and a spatiotemporal data sequence is constructed based on timestamps and location identifiers. The spatiotemporal data sequence undergoes timestamp alignment, outlier removal, noise reduction, and standardization, and data is aggregated according to a preset time window to form a multidimensional time series sample. Multi-source data is fused based on timestamps and location identifiers to construct a temporal feature vector containing operating parameters and corrosion characteristics. The temporal feature vector is then input into a convolutional neural network. Feature extraction is performed in a combined model of network and long short-term memory network to obtain corrosion evolution feature representation; corrosion type identification is performed on the corrosion evolution feature representation based on deep neural network model, and future corrosion rate is predicted based on time series prediction model; the remaining service life is calculated based on the current wall thickness value, predicted corrosion rate and preset safety threshold, and the prediction results are constrained or corrected using corrosion rate mechanism formula; early warning assessment is performed based on corrosion type identification results, corrosion rate prediction results and remaining service life, and the deep learning model is incrementally trained based on measured data after maintenance to achieve adaptive model update.

[0009] Furthermore, the multidimensional time series samples are constructed using a sliding time window, which contains data from multiple consecutive time steps.

[0010] Furthermore, the time-series feature vector includes wall thickness, temperature, pressure, vibration amplitude, vibration frequency, medium sulfur content, and flow velocity parameters.

[0011] Preferably, the corrosion rate is calculated according to the following formula: v(t) = (D prev - D curr ) / Δt, Where v(t) represents the corrosion rate at time t, D prev D represents the wall thickness value at the previous detection time. curr Δt represents the wall thickness value at the current detection time, and Δt represents the time interval between the previous detection time and the current detection time.

[0012] Preferably, the remaining service life is calculated according to the following formula: RUL = (Dcurr - D threshold ) / v pred , Where RUL represents the remaining useful life, D curr D represents the wall thickness value at the current detection moment. threshold Indicates the critical wall thickness value for safety, v pred This indicates the predicted corrosion rate.

[0013] Optionally, it also includes comprehensive confidence calculation and data augmentation, wherein the comprehensive confidence is calculated as follows: C onf = α × Acc + β × (1 - RMSE / Dcurr), Among them, C onf The overall confidence level is represented by α, the weighting coefficient of the corrosion type identification result is represented by β, the weighting coefficient of the corrosion rate prediction result is represented by β, and α + β = 1, Acc represents the corrosion type identification accuracy, and RMSE represents the root mean square error of corrosion rate prediction.

[0014] In one embodiment of the present invention, an intelligent diagnostic system for corrosion faults in atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints is also provided. This system includes a data acquisition unit, an edge processing unit, a cloud analysis unit, an early warning and visualization unit, and a model update unit. The data acquisition unit collects wall thickness, temperature, pressure, vibration, and medium composition data at multiple spatial locations within the atmospheric pressure tower and constructs a spatiotemporal data sequence with timestamps and location identifiers. The edge processing unit is communicatively connected to the data acquisition unit and performs timestamp alignment, outlier removal, noise reduction, and standardization on the spatiotemporal data sequence, and processes it according to a preset time window. A multidimensional time series sample is generated. The cloud analysis unit is communicatively connected to the edge processing unit and is used to perform fusion modeling on the multidimensional time series sample, extract corrosion evolution features, perform corrosion type identification, corrosion rate prediction and remaining service life calculation, and constrain or correct the prediction results based on the corrosion mechanism formula. The early warning and visualization unit is communicatively connected to the cloud analysis unit and is used to trigger early warning based on corrosion rate, remaining service life and diagnostic results, and to perform spatial location visualization. The model update unit is communicatively connected to the cloud analysis unit and is used to receive maintenance data and perform incremental training and parameter optimization on the model.

[0015] Furthermore, the data acquisition unit includes an ultrasonic thickness sensor, a temperature sensor, a pressure sensor, a vibration sensor, a near-infrared spectrometer, and an interface for reading process parameters from the distributed control system.

[0016] Preferably, the edge processing unit adopts an industrial edge computing gateway and is configured with local early warning rules to directly trigger an on-site alarm without going through the cloud when the corrosion rate exceeds a preset threshold.

[0017] Optionally, the cloud-based analysis unit further includes a model interpretability module for calculating the contribution of input features to the diagnostic results based on Shapley values ​​or the Locally Interpretable Model Irrelevant Interpretation (LIME) algorithm.

[0018] Based on the above technical solutions, the present invention provides an intelligent diagnostic method and system for corrosion faults in atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints. This method constructs a multi-source spatiotemporal data sequence containing timestamps and spatial location identifiers, performs unified preprocessing and fusion modeling on heterogeneous data from multiple sources such as wall thickness, temperature, pressure, vibration, and medium composition, extracts the temporal features of corrosion evolution using a model combining convolutional neural networks and long short-term memory networks, accurately identifies corrosion types through deep neural networks, and dynamically predicts corrosion rates using a time series prediction model. At the same time, a corrosion rate mechanism formula is introduced to constrain or correct the prediction results, thereby achieving quantitative calculation of the remaining service life.

[0019] Furthermore, by comprehensively analyzing the corrosion type identification results, corrosion rate prediction results, and remaining service life assessment results, a multi-dimensional assessment and graded early warning of the corrosion status of the atmospheric pressure tower can be achieved. Combined with digital twin technology, spatial visualization of corrosion location can be realized, making the equipment operation status more intuitive and traceable. At the same time, by introducing a maintenance measurement data feedback mechanism, the deep learning model is incrementally trained and continuously optimized, constructing a closed-loop self-learning system of "data acquisition - intelligent diagnosis - prediction assessment - feedback update".

[0020] Compared with the prior art, the present invention has at least the following beneficial effects: (1) Achieve continuous and global monitoring: Through multi-source sensors and spatiotemporal data modeling, achieve continuous online monitoring of key parts of atmospheric pressure tower, and overcome the problems of data discontinuity and delayed detection of corrosion anomalies in traditional manual inspection and offline detection; (2) Improve the accuracy of corrosion diagnosis: By using deep learning models to fuse and analyze multi-source data, it is possible to effectively identify various types of corrosion, reduce reliance on human experience, and improve the objectivity and accuracy of diagnostic results; (3) Achieve dynamic prediction of corrosion evolution: By combining time series models and mechanistic constraints, achieve dynamic prediction of corrosion rate and quantitative assessment of remaining service life, providing a forward-looking decision-making basis for equipment operation; (4) Enhance model reliability and interpretability: Constrain the model prediction results by using mechanistic formulas to improve the physical rationality of the prediction results. At the same time, combine interpretability analysis methods to make the diagnostic results traceable. (5) Construct a closed-loop optimization mechanism: By introducing maintenance data feedback and model incremental training mechanism, the model can continuously learn and improve its performance, so that the system can adapt to complex working conditions. (6) Achieve intelligent early warning and visual management: Through a multi-level early warning mechanism and three-dimensional digital twin visualization, the corrosion risk is presented intuitively and actively pushed, thereby improving the safety of equipment operation; (7) Reduce maintenance costs and improve equipment availability: By replacing the traditional periodic maintenance mode with predictive maintenance, the risk of unplanned downtime is reduced, maintenance strategies are optimized, thereby reducing maintenance costs and improving equipment operating efficiency.

[0021] In summary, this invention combines multi-source data fusion, deep learning modeling, and corrosion mechanism constraints to form a complete intelligent diagnosis and prediction system for corrosion of atmospheric pressure towers. It can not only realize real-time perception and accurate identification of corrosion status, but also predict and evaluate the future operating status of equipment, which has significant engineering application value and promotion prospects. Attached Figure Description

[0022] Figure 1This is a flowchart illustrating an intelligent diagnostic method for corrosion faults in atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints, provided in an embodiment of the present invention. The process includes steps such as data acquisition, data preprocessing, multi-source data fusion, feature extraction, corrosion type identification, corrosion rate prediction, remaining service life calculation, early warning assessment, and model updating. Figure 2 This is a schematic diagram of the structure of an intelligent diagnostic system for corrosion faults in atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints, provided in an embodiment of the present invention. The system includes a data acquisition unit, an edge processing unit, a cloud analysis unit, an early warning and visualization unit, and a model update unit. The units are connected by communication to achieve data interaction and collaborative work.

[0023] Explanation of reference numerals in the attached figures: 100 represents the data acquisition unit, 200 represents the edge processing unit, 300 represents the cloud analysis unit, 400 represents the early warning and visualization unit, and 500 represents the model update unit. Detailed Implementation

[0024] To make the technical solution of the present invention clearer and more complete, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that various modifications or substitutions can be made to the following embodiments without departing from the spirit and scope of the present invention.

[0025] I. General Description To make the technical solution of the present invention clearer and more complete, the present invention will be further described in detail below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the various embodiments and technical features in the embodiments described in this specification can be combined with each other.

[0026] In one embodiment of the present invention, such as Figure 1 As shown, the overall flowchart of an intelligent diagnosis method for corrosion faults in atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints is presented. This method mainly includes the following steps: First, data on wall thickness, temperature, pressure, vibration, and medium composition were collected at multiple spatial locations in the atmospheric pressure tower, and a spatiotemporal data sequence was constructed based on timestamps and spatial location identifiers. Then, the spatiotemporal data sequence is preprocessed, including timestamp alignment, outlier removal, noise reduction and standardization, and the data is aggregated according to a preset time window to form a multidimensional time series sample. Furthermore, the multi-source data are fused to construct a time-series feature vector that includes operating parameters and corrosion characteristics; Subsequently, the temporal feature vector is input into a combined model of convolutional neural network and long short-term memory network for feature extraction to obtain feature representations characterizing the corrosion evolution process, and corrosion type identification is achieved based on a deep neural network model. Based on this, a time series prediction model is used to predict the future corrosion rate, and the predicted corrosion rate v is obtained. pred And combined with the wall thickness value D at the current detection time curr The wall thickness Dprev at the previous detection time and the time interval Δt are used to calculate the corrosion rate using the formula v(t) = (Dprev / Δt). prev - D curr The prediction results are constrained or corrected by Δt. Furthermore, based on the corrected corrosion rate and the critical safety wall thickness D... threshold According to the remaining useful life calculation formula RUL = (Dcurr - D) threshold ) / v pred Calculate the remaining useful life of the equipment; Then, based on the corrosion type identification results, corrosion rate prediction results, and remaining service life, the equipment operating status is comprehensively evaluated and an early warning is triggered. Finally, the measured data obtained after equipment maintenance is used as feedback samples to incrementally train the deep learning model, thereby updating the model parameters and forming a continuously optimized closed-loop diagnostic mechanism.

[0027] In one embodiment of the present invention, such as Figure 2 The diagram shows a schematic representation of an intelligent diagnostic system for corrosion faults in atmospheric pressure towers, used to implement the aforementioned method. The system includes: The system includes a data acquisition unit 100, an edge processing unit 200, a cloud analysis unit 300, an early warning and visualization unit 400, and a model update unit 500. The data acquisition unit 100 is used to collect wall thickness, temperature, pressure, vibration and medium composition data at multiple spatial locations in the atmospheric pressure tower, and to construct a spatiotemporal data sequence with timestamps and spatial location identifiers. The edge processing unit 200 is communicatively connected to the data acquisition unit 100 and is used to perform timestamp alignment, outlier removal, noise reduction and standardization on the spatiotemporal data sequence, and generate multidimensional time series samples according to a preset time window. The cloud analysis unit 300 is communicatively connected to the edge processing unit 200, and is used to perform fusion modeling of multi-dimensional time series samples, extract corrosion evolution features, and perform corrosion type identification, corrosion rate prediction, and remaining service life calculation, based on the corrosion rate formula v(t) = (D prev - D currThe formula for RUL = (D / Δt and remaining useful life) is: curr - D threshold ) / v pred Constrain or correct the model's prediction results; The early warning and visualization unit 400 is communicatively connected to the cloud analysis unit 300, and is used to trigger early warnings based on corrosion rate, remaining service life and diagnostic results, and to spatially visualize the corrosion status of each measuring point of the atmospheric pressure tower. The model update unit 500 is communicatively connected to the cloud analysis unit 300, and is used to receive the measured data after maintenance, and to perform incremental training and parameter optimization on the deep learning model in order to improve the diagnostic accuracy and adaptability of the model.

[0028] Through the coordinated use of the above methods and systems, real-time monitoring, intelligent diagnosis, dynamic prediction, and closed-loop optimization of the corrosion status of atmospheric pressure towers were achieved.

[0029] II. Example 1: Data Acquisition and System Deployment In this embodiment, corresponding Figure 1 Step S1 and Figure 2 The data acquisition unit 100 in the middle performs multi-source data acquisition and system deployment for the atmospheric pressure tower.

[0030] In one embodiment of the present invention, the data acquisition unit 100 is deployed at multiple key spatial locations in the atmospheric pressure tower to acquire multi-source data reflecting the corrosion status and operating conditions of the equipment. These key spatial locations include, but are not limited to, corrosion-sensitive areas such as the tower wall, tray support locations, downcomer area, feed section, top evaporation line, and side feed outlet.

[0031] Furthermore, the data acquisition unit 100 includes various types of sensors and detection devices, specifically including: Ultrasonic thickness sensor is used to measure the wall thickness of various measuring points in an atmospheric pressure tower in real time. Temperature sensor, used to collect temperature data at the corresponding location; Pressure sensors are used to collect operating pressure data of equipment; Vibration sensors are used to acquire vibration amplitude and vibration characteristic data of equipment; A media composition acquisition module or near-infrared spectroscopy analyzer is used to detect corrosion-related components in the media, such as sulfur content and acid value.

[0032] In this embodiment, the ultrasonic thickness sensor is preferably a high-temperature type thickness sensor to adapt to the high-temperature operating environment of the atmospheric pressure tower; the temperature sensor is preferably a thermocouple or a resistance temperature detector; the pressure sensor is preferably an industrial-grade pressure transmitter; and the vibration sensor is preferably a high-temperature acceleration sensor.

[0033] Furthermore, each sensor collects the corresponding physical quantity according to a preset sampling frequency, wherein: Wall thickness data can be collected at a frequency of minutes or higher; Temperature, pressure, and vibration data can be acquired at the second level or higher frequency. Media composition data can be collected periodically on a minute-by-minute or lower frequency.

[0034] Preferably, the sampling frequency of different types of data is configured according to the actual working conditions to reduce system load while ensuring data accuracy.

[0035] In this embodiment, the data acquisition unit 100 interacts with the edge processing unit 200 through an industrial communication network. The communication methods include, but are not limited to, industrial Ethernet, RS-485 bus, or wireless communication network.

[0036] Furthermore, to ensure data consistency and fusionability, a standardized information identifier is attached to each piece of collected data during the data collection process, including: Timestamp information is used to identify the specific time of data collection; Spatial location identifiers are used to identify the location of the measurement points corresponding to the data.

[0037] The timestamp is preferably generated based on a unified clock source, such as through GPS or a network time protocol, to ensure time consistency among multiple data sources.

[0038] By using the above method, multi-source data from different sensors are uniformly organized into a spatiotemporal data sequence containing time and spatial dimensions, providing a foundation for subsequent data preprocessing, multi-source fusion, and intelligent analysis.

[0039] In addition, in this embodiment, process parameter data related to the operation of the atmospheric distillation column can also be obtained from the distributed control system (DCS), including but not limited to: throughput, reflux ratio, extraction volume of each side stream, temperature and pressure at the top and bottom of the column.

[0040] The process parameters, as supplementary data, together with the data collected by the sensors, constitute multi-source data input, thereby improving the corrosion diagnosis model's ability to characterize complex working conditions.

[0041] Through the above data acquisition and system deployment scheme, a comprehensive understanding of the corrosion status and operating conditions of key parts of the atmospheric pressure tower was achieved, providing a reliable data foundation for subsequent intelligent diagnosis and predictive analysis.

[0042] III. Example 2: Data Preprocessing and Time Series Construction In this embodiment, corresponding Figure 1 Step S2 and Figure 2 The edge processing unit 200 in the middle performs preprocessing and time series construction on the multi-source spatiotemporal data acquired by the data acquisition unit 100.

[0043] In one embodiment of the present invention, the edge processing unit 200 is deployed on the field side of the atmospheric pressure tower and is communicatively connected to the data acquisition unit 100. It is used to process the acquired raw spatiotemporal data sequence in real time to improve data quality and reduce the computational burden of subsequent cloud analysis.

[0044] Furthermore, the edge processing unit 200 performs the following preprocessing steps on the raw data in sequence: First, timestamp alignment is performed.

[0045] Because different sensors have different sampling frequencies and communication delays, the original data may be offset in the time dimension. Therefore, by using a unified time base to resample and align the multi-source data, different types of data can correspond at the same time scale, thus forming a synchronized data sequence.

[0046] Preferably, the time alignment can be achieved using interpolation or window matching methods to ensure data integrity.

[0047] Secondly, outlier removal is performed.

[0048] In industrial environments, factors such as electromagnetic interference and sensor jitter can cause abrupt changes or invalid data points in the collected data. In this embodiment, a statistical anomaly detection method is used to filter the data.

[0049] Preferably, the outlier removal is based on the 3σ criterion, which means that data points exceeding the mean ± 3 times the standard deviation are identified as outliers and removed or corrected.

[0050] Furthermore, for the missing data after removal, linear interpolation or nearest neighbor imputation can be used to complete the data in order to maintain the continuity of the time series.

[0051] Then, noise reduction processing is performed.

[0052] Since vibration data and some process parameter signals may contain high-frequency noise, this embodiment preferably uses wavelet transform to denoise the data in order to retain the main characteristic components of the signal while suppressing random noise interference.

[0053] Furthermore, different noise reduction strategies can be selected for different types of data based on their signal characteristics to improve processing performance.

[0054] Next, standardization processes will be implemented.

[0055] Since various types of data differ in their dimensions and numerical ranges, in order to improve the stability of model training, it is necessary to normalize or standardize the data so that the features of each dimension are in similar numerical ranges.

[0056] Preferably, the standardization method can be min-max normalization or Z-score standardization.

[0057] After completing the above preprocessing, time window aggregation and time series construction are further performed.

[0058] Specifically, the preprocessed data is segmented and aggregated according to a preset time window to form a multidimensional time series sample. The time window can be set according to actual application requirements, for example: Short time windows are used to capture instantaneous changes. Long-term windows are used to reflect the long-term evolution trend of corrosion.

[0059] Preferably, the time window is constructed using a sliding window method, that is, the window slides along the time axis at fixed steps, thereby generating multiple overlapping time series samples and improving data utilization.

[0060] Furthermore, within each time window, statistical features can be extracted from the raw data, for example: Maximum value Minimum value average value Standard deviation This reduces data dimensionality while retaining key feature information.

[0061] Through the above processing, we finally obtain multidimensional time series samples with uniform structure, each sample containing: Time dimension information Spatial location identifier Multi-source feature data This provides standardized input for subsequent multi-source data fusion, feature extraction, and deep learning model training.

[0062] Furthermore, in this embodiment, the edge processing unit 200 can also perform local rapid response processing according to preset rules. For example, when an abnormal increase in corrosion rate or abnormal change in wall thickness is detected, a local early warning can be directly triggered, thereby improving the system response speed.

[0063] Through the above data preprocessing and time series construction process, the data quality and consistency are effectively improved, the fusionability between multi-source data is enhanced, and a reliable data foundation is provided for subsequent corrosion diagnosis and predictive analysis based on deep learning.

[0064] IV. Example 3: Multi-source data fusion and feature construction In this embodiment, corresponding Figure 1 In step S3, the multi-source data fusion and feature construction are performed on the multi-dimensional time series samples processed by the edge processing unit 200 to form a time series feature vector suitable for the input of the deep learning model.

[0065] In one embodiment of the present invention, the multi-source data fusion is based on timestamps and spatial location identifiers. Specifically, multidimensional time-series data from different sensors and data sources are aligned according to a unified time scale, and spatial location identifiers are used as indexes to integrate various types of data from the same measurement point within the same time window, thereby constructing a unified data representation.

[0066] Furthermore, for any spatial location i, within the time window Internally, the corresponding multi-source data are fused into a single multi-dimensional feature vector. Its form can be expressed as:

[0067] in: This represents the wall thickness at position i at time t; Represents temperature data; Indicates pressure data; Indicates the vibration amplitude; Indicates the dominant frequency or frequency domain characteristics of vibration; Indicates the composition parameters of the medium (such as sulfur content or acid value); This indicates the flow velocity or the flow characteristics calculated from process parameters. The above method transforms heterogeneous and scattered multi-source data into a unified structured feature vector.

[0068] Furthermore, to reflect the temporal evolution characteristics of the corrosion process, feature vectors from multiple consecutive time steps are concatenated to construct a temporal feature input matrix. For example, for a time window of length L, the following input data structure can be constructed:

[0069] Thus forming a dimension of The time-series feature matrix, where The dimension of a single feature vector.

[0070] Preferably, the time window length L can be set according to the actual working conditions to balance the ability to express short-term fluctuation characteristics and long-term corrosion trends.

[0071] In this embodiment, to improve the model's ability to represent complex working conditions, further feature expansion and construction can be performed, including but not limited to: 1. Construction of statistical features Extract statistical features, such as mean, standard deviation, and extreme values, within the time window to enhance the description of the data distribution characteristics; 2. Construction of Coupling Features We constructed coupling characteristics between variables such as temperature and pressure, vibration and flow rate to reflect corrosion behavior under the combined effect of multiple factors; 3. Construction of Trend Characteristics By calculating indicators such as wall thickness change rate and temperature change gradient, the model's ability to perceive corrosion evolution trends is enhanced. 4. Normalized Feature Representation The constructed feature vectors are subjected to uniform scaling to improve the stability of model training.

[0072] Furthermore, in this embodiment, the corrosion rate can be calculated based on historical detection data as a supervised learning label, and the calculation method is as follows: v(t) = (D prev - D curr ) / Δt Among them, D prev D represents the wall thickness value at the previous detection time. curr Δt represents the wall thickness at the current detection moment, and Δt represents the time interval.

[0073] The corrosion rate is used as label data and combined with the corresponding temporal feature input to form training samples, which are then used to train the subsequent corrosion rate prediction model.

[0074] In addition, in this embodiment, corrosion types can be labeled based on historical maintenance records. For example, corrosion states can be divided into: normal state, uniform corrosion, pitting corrosion, erosion corrosion, and stress corrosion cracking, thereby constructing label data for training classification models.

[0075] Through the above-mentioned multi-source data fusion and feature construction process, the original industrial data is transformed into structured, time-seriesd, and physically meaningful model input data. This not only improves the efficiency of data utilization but also enhances the model's ability to express corrosion mechanisms, providing key support for subsequent feature extraction and intelligent diagnosis.

[0076] V. Example 4: Feature Extraction and Fault Identification In this embodiment, corresponding Figure 1 Steps S4 and S5 in the process and Figure 2The cloud-based analysis unit 300 performs feature extraction and corrosion fault identification on the multi-source fused time-series feature data.

[0077] In one embodiment of the present invention, the cloud analysis unit 300 is used to receive multi-dimensional time series samples transmitted by the edge processing unit 200, and analyze and process the data based on a deep learning model to extract corrosion evolution features and realize corrosion type identification.

[0078] (I) Feature Extraction Process In this embodiment, the temporal feature input matrix constructed in Embodiment 3 is used as the model input and input into the feature extraction model that combines convolutional neural network and long short-term memory network.

[0079] Furthermore, the feature extraction model includes: 1. One-dimensional convolutional neural network (1D-CNN) layer The 1D-CNN is used for local feature extraction of time series data, and its functions include: Capture local variation patterns of multi-source data such as wall thickness, temperature, and pressure over time; Extract local feature information reflecting changes in the early stages of corrosion; Preferably, the 1D-CNN layer includes: Multiple convolutional layers, with convolutional kernels sliding along the time dimension; The activation function used is the ReLU function; Different convolutional kernel sizes can be configured to extract features at different time scales; Furthermore, by using a multi-layer convolutional structure, we can achieve layer-by-layer feature abstraction of the input data.

[0080] 2. Pooling layer After the convolutional layer, a pooling layer can be set for: Reduce feature dimensionality; Suppress noise interference; Improve the model's generalization ability; Preferably, the pooling layer employs max pooling.

[0081] 3. Long Short-Term Memory (LSTM) Layer In this embodiment, the features extracted by convolution are input into an LSTM network to model the time dependency of the erosion process.

[0082] Specifically, the LSTM layer is used for: Capture long-term trends during the corrosion process; Characterizing the dynamic evolution of corrosion rate; Establish relationships between different time steps; Preferably, the LSTM layer outputs a hidden state vector h_t, which serves as a high-dimensional feature representation for subsequent classification and prediction.

[0083] By combining 1D-CNN and LSTM, the system can extract features from multi-source temporal data from "local features" to "global temporal features" layer by layer, thereby obtaining feature representations that characterize the erosion evolution process.

[0084] (II) Corrosion Fault Type Identification In this embodiment, the feature representation output from the feature extraction stage is input into a deep neural network (DNN) classification model to achieve corrosion fault type identification.

[0085] Furthermore, the DNN model includes: Multiple fully connected layers; Each layer uses a non-linear activation function (preferably the ReLU function); The output layer uses the Softmax function for multi-class classification output; Specifically, the DNN model outputs the probability distribution of each corrosion type based on the input features, and selects the one with the highest probability as the final recognition result.

[0086] In this embodiment, the corrosion types include, but are not limited to, the following categories: Normal state; Uniform corrosion; Pitting; Erosion and corrosion; Stress corrosion cracking.

[0087] (III) Model Training and Optimization In this embodiment, the feature extraction model and classification model can be trained using historical data.

[0088] Specifically: The input data is the multidimensional temporal feature matrix constructed in Example 3; The label data includes corrosion type labels and labels derived from formulas. v(t) = (D prev - D curr ) / Δt The calculated corrosion rate label; Preferably, the classification model is trained using the cross-entropy loss function and the model parameters are updated using the backpropagation algorithm.

[0089] Furthermore, to prevent model overfitting, the following can be introduced during training: Dropout mechanism; Regularization term; Batch Normalization; (iv) Interpretability Analysis In this embodiment, the cloud analysis unit 300 may further include a model interpretability module for interpreting the corrosion type identification results.

[0090] Specifically, a method based on Shapley values ​​or... Locally Interpretable Model-Independent Interpretation Algorithm (LIME) The importance of the input features is evaluated to obtain the contribution of each feature to the current diagnostic result.

[0091] The above methods can be used to identify the key factors that have the greatest impact on corrosion, such as: Sulfur content of the medium; Temperature fluctuations; Abnormal vibration; This provides a basis for decision-making for equipment maintenance personnel.

[0092] (v) Technical effects of this embodiment Through the above feature extraction and fault identification process, the following was achieved: Deep feature mining of multi-source data; Automated identification of corrosion types; Accurate modeling of corrosion behavior under complex working conditions; It also provides high-quality input features for subsequent corrosion rate prediction and remaining service life assessment.

[0093] VI. Example 5: Corrosion Rate Prediction and Lifetime Assessment In this embodiment, corresponding Figure 1 Steps S6 and S7 in the process and Figure 2 The cloud-based analysis unit 300 in the system performs corrosion rate prediction and remaining service life assessment on the corrosion evolution characteristics after feature extraction.

[0094] (a) Corrosion rate prediction In one embodiment of the present invention, the corrosion evolution characteristics obtained in Example 4 are input into a time series prediction model to predict the corrosion rate in future time periods, thereby obtaining the predicted corrosion rate v. pred .

[0095] Furthermore, the time series prediction model can employ: Temporal Convolutional Network (TCN); Sequence-to-sequence model (Seq2Seq); Or other deep learning models capable of modeling time dependencies; Used to study the dynamic changes in the corrosion process.

[0096] Specifically, the model takes the time-series feature vector output from the feature extraction stage as input, and outputs a predicted corrosion rate value for one or more future time steps by learning from historical time series data.

[0097] Preferably, the prediction result may include: Short-term corrosion rate prediction (e.g., for the next few hours or days); Medium- to long-term corrosion rate prediction (e.g., for the next few weeks or months); To meet the needs of different operation and maintenance scenarios.

[0098] (II) Corrosion rate constraint based on mechanism formula In this embodiment, to improve the physical rationality of the prediction results, a corrosion rate calculation formula based on wall thickness variation is introduced to constrain or correct the model output.

[0099] The corrosion rate is calculated according to the following formula: v(t) = (D prev - D curr ) / Δt in: D prev This indicates the wall thickness value at the previous detection time. D curr This indicates the wall thickness value at the current detection moment; Δt represents the time interval; Furthermore, comparing the corrosion rate calculated by the above formula with the model prediction value v_pred can be used for: Correct the prediction results; As a supervisory label for model training; Construct the constraint terms in the loss function; This makes the prediction results more consistent with the actual corrosion mechanism.

[0100] (III) Remaining useful life assessment In this embodiment, after obtaining the predicted corrosion rate v pred Then, the remaining service life of the equipment is further assessed.

[0101] The remaining useful life is calculated according to the following formula: RUL = (D curr - D threshold ) / v pred in: D curr This indicates the wall thickness value at the current detection moment; D threshold Indicates the critical wall thickness value for safety; v pred This indicates the predicted corrosion rate; The above calculations can be used to determine the remaining safe operating time of the equipment under the current corrosion trend.

[0102] Furthermore, the safety critical wall thickness D threshold Based on: Equipment design standards; Safety regulations; Historical operational experience; Configure the settings.

[0103] (iv) Dynamic update mechanism In this embodiment, the remaining useful life assessment is a dynamic calculation process.

[0104] Specifically, as new test data is continuously input: D curr Updated in real time; v pred Continuous forecasting; This enables dynamic updates to the Remaining Service Life (RUL), allowing the assessment results to reflect the current true condition of the equipment.

[0105] (V) Multi-timescale prediction and fusion Furthermore, in this embodiment, corrosion rates at different time scales can be predicted simultaneously, for example: Short-term forecasts (used for rapid early warning); Long-term forecasts (used for decision maintenance); It can also perform weighted fusion of prediction results at different time scales to improve prediction stability and reliability.

[0106] (vi) Application of prediction results In this embodiment, the corrosion rate prediction results and the remaining service life assessment results can be used for: As input for subsequent early warning assessment; Support equipment maintenance strategy formulation; Optimize maintenance schedules; Furthermore, by combining the corrosion type identification results, a comprehensive judgment on the risk level under different corrosion modes can be achieved.

[0107] This embodiment achieves the following by combining a deep learning prediction model with a corrosion mechanism formula: High-precision prediction of corrosion rate; Quantitative assessment of remaining useful life; Physical consistency constraints on model prediction results; This effectively improves the reliability and engineering applicability of the corrosion diagnosis system.

[0108] VII. Example 6: Early Warning and Visualization In this embodiment, corresponding Figure 1 Step S8 and Figure 2 The early warning and visualization unit 400 in the system comprehensively analyzes the corrosion rate prediction results, remaining service life assessment results, and corrosion type identification results, and realizes early warning and visualization display of equipment operating status.

[0109] (I) Early warning and assessment mechanism In one embodiment of the present invention, the early warning and visualization unit 400 is communicatively connected to the cloud analysis unit 300 and is used to receive the following input data: Corrosion type identification results; Corrosion rate prediction results v pred ; Remaining useful life (RUL); Current wall thickness value D curr ; Furthermore, based on the above data, a comprehensive assessment of the corrosion status at each measuring point of the atmospheric pressure tower was conducted.

[0110] Preferably, the early warning assessment is based on preset threshold rules, and the thresholds include, but are not limited to: Corrosion rate threshold; Remaining useful life threshold; Wall thickness safety threshold D threshold ; The operating status of the equipment is determined by comparing the actual monitored values ​​with the threshold values.

[0111] (II) Multi-level early warning strategy In this embodiment, to improve the accuracy and operability of early warning, corrosion risk is divided into multiple levels, for example: Normal condition: Low corrosion rate, sufficient remaining service life; Note: The corrosion rate has increased, but has not reached a dangerous level; Warning status: Corrosion rate is high or remaining service life is close to the threshold; Alarm status: Corrosion rate significantly exceeds the limit or remaining service life is insufficient; Furthermore, different levels correspond to different processing strategies, for example: Normal state: Continuous monitoring; Note the status: Increase the detection frequency; Warning status: Arrange maintenance plans; Alarm status: Immediately take shutdown or emergency measures; The aforementioned grading mechanism enables refined management of corrosion risks.

[0112] (III) Integration of Comprehensive Assessment and Confidence Level In this embodiment, a comprehensive confidence level can also be introduced to correct the warning results.

[0113] Specifically, the overall confidence level is calculated as follows: Conf = α × Acc + β × (1 - RMSE / Dcurr) By introducing a confidence level parameter, the reliability of diagnostic results can be assessed. When the confidence level is low, the warning level can be reduced or manual review can be triggered, thereby avoiding false alarms or missed alarms.

[0114] (iv) Visualization of digital twins In this embodiment, the early warning and visualization unit 400 is also used to realize the spatial visualization display of the corrosion state of the atmospheric pressure tower.

[0115] Specifically, a three-dimensional digital twin model was established based on the atmospheric pressure tower structure, and the corrosion status of each measuring point was mapped to the corresponding spatial location.

[0116] Furthermore, different colors can be used to identify each measuring point, for example: Green indicates the normal state; Yellow indicates a state of alertness; Orange indicates a warning status; Red indicates an alarm status; The above methods can provide a direct visual representation of the corrosion distribution.

[0117] (v) Interactive information display Furthermore, in this embodiment, the digital twin model supports interactive operation. When a user selects a measurement point, detailed information about that measurement point can be displayed, including: Current wall thickness value D curr ; Corrosion rate prediction value v pred ; Remaining useful life (RUL); Corrosion type; Key influencing factors (combined with interpretability analysis results); This provides comprehensive data support for operations and maintenance personnel.

[0118] (vi) Early warning information push mechanism In this embodiment, the warning result can also be pushed to the terminals of relevant personnel via the communication system, including: Engineer's mobile terminal; Monitoring platform; Enterprise operation and maintenance management system; Preferably, the push content includes: Warning level; Equipment location; Risk description; Recommended measures; This enables remote monitoring and rapid response.

[0119] (vii) Technical effects of this embodiment Through the aforementioned early warning and visualization mechanisms, the following was achieved: Real-time assessment and hierarchical management of corrosion risks; A visual representation of the corroded state in the space; Timely transmission of early warning information; Intelligent support for operation and maintenance decision-making; This significantly improves the practicality and engineering application value of the atmospheric pressure tower corrosion monitoring system.

[0120] VIII. Example 7: Model Update and Closed-Loop Optimization In this embodiment, corresponding Figure 1 Step S9 and Figure 2 The model update unit 500 in the system introduces maintenance feedback data to perform incremental training and parameter optimization of the deep learning model, thereby achieving closed-loop self-learning and continuous performance improvement of the system.

[0121] (a) Feedback Data Acquisition Mechanism In one embodiment of the present invention, the model update unit 500 is used to acquire actual maintenance data during equipment operation, the maintenance data including but not limited to: Measured wall thickness (used to update D) curr ); Actual corrosion depth and corrosion rate; Corrosion type (e.g., pitting corrosion, uniform corrosion, etc.); Non-destructive testing results (such as X-ray or ultrasonic testing results); Furthermore, the above data can be obtained through the following methods: On-site inspection records of maintenance personnel; Data entry into the enterprise operations and maintenance system; Automatic detection equipment collects data; This results in the formation of truly labeled samples.

[0122] (II) Data Labeling and Sample Construction In this embodiment, the maintenance data is matched with historical monitoring data, and labeled training samples are constructed.

[0123] Specifically: Associate the time-series characteristic data within a historical time period with the corresponding maintenance results; Label the corrosion type according to the inspection results; Based on wall thickness variation data, according to the formula v(t) = (D prev - D curr ) / Δt Calculate the actual corrosion rate label; The above methods are used to construct a high-quality supervised learning dataset.

[0124] (III) Incremental Model Training Mechanism In this embodiment, the model update unit 500 performs incremental training on the existing model based on the new samples, without having to retrain the entire model.

[0125] Furthermore, the incremental training process includes: Use the current model parameters as the initial weights; The newly added maintenance data was used as training data; Fine-tuning the model using a smaller learning rate; Preferably, the learning rate is lower than that in the initial training phase to avoid drastic fluctuations in the model.

[0126] Furthermore, incremental training can employ a periodic triggering mechanism, for example: Updated daily; Updated weekly; Or it may be triggered when the number of new samples reaches a preset quantity; This enables continuous optimization of the model.

[0127] (iv) Model performance optimization strategy In this embodiment, to improve the model training effect, the following optimization strategies can be combined: Regularization methods are introduced to prevent overfitting; Use early stopping strategies to avoid overtraining; Balancing samples from different categories improves the stability of the classification model; Furthermore, for a small number of corrosion samples, supplementary samples can be generated by combining data augmentation methods (such as generative adversarial networks) to improve the model's generalization ability.

[0128] (v) Closed-loop optimization mechanism In this embodiment, a complete closed-loop system is constructed through the process of "data acquisition - intelligent diagnosis - predictive evaluation - maintenance feedback - model update".

[0129] The specific process is as follows: 1. Data acquisition unit 100 acquires real-time data; 2. The edge processing unit 200 performs data preprocessing; 3. The cloud-based analytics unit performs diagnostics and predictions. 4. The early warning and visualization unit 400 outputs early warning results; 5. Obtain accurate data after equipment maintenance; 6. Model update unit 500 performs incremental training on the model; Through the aforementioned closed-loop mechanism, the model can continuously adapt to new changes in operating conditions.

[0130] (vi) Model performance improvement effect In this embodiment, by continuously introducing real maintenance data, the model performance can be gradually improved, for example: The accuracy of corrosion type identification has been improved; The error in corrosion rate prediction has been reduced; The remaining useful life assessment is more accurate; Furthermore, as the scale of data increases, the model's ability to identify complex corrosion patterns continues to improve.

[0131] Through the above model update and closed-loop optimization mechanism, the following was achieved: The model is shifting from "static training" to "dynamic learning"; Improve the model's adaptability to different working conditions; Reduce model error and improve prediction reliability; Build a sustainable and optimized intelligent diagnostic system; This significantly enhances the system's long-term application value and engineering practicality.

[0132] IX. Example 8: Comparative Experiment In this embodiment, to verify the effectiveness of the intelligent diagnosis method and system for corrosion faults of atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints described in this invention, a comparative experimental analysis is conducted between the method of this invention and existing technical methods.

[0133] (a) Sources of experimental data In this embodiment, historical operating data of an atmospheric distillation unit in an oil refinery were selected as experimental data. The data sources include: Wall thickness monitoring data from multiple measuring points; Data on temperature, pressure, vibration, and media composition; Actual maintenance records and corrosion test results; The data covers multiple operating cycles and includes different corrosion types and various operating condition changes.

[0134] (II) Setting up the comparison method In this embodiment, the following method is selected as a comparison scheme: 1. Traditional empirical methods Corrosion assessment is based on human experience and simple threshold rules; 2. Single machine learning method Corrosion classification is performed using traditional machine learning models (such as support vector machines or random forests); 3. Pure Deep Learning Methods Deep learning models (such as CNN or LSTM) are used, but without introducing erosion mechanism constraints; 4. The method of the present invention It employs multi-source data fusion, CNN-LSTM feature extraction, time series prediction and mechanism constraint fusion, combined with an incremental model update mechanism.

[0135] (III) Evaluation Indicators In this embodiment, the following metrics are used to evaluate each method: Corrosion type identification accuracy (Acc); Corrosion rate prediction error (RMSE); Remaining useful life prediction error; Early warning time; in: Accuracy is used to evaluate classification performance; RMSE is used to evaluate prediction accuracy; Early warning lead time is used to assess the system's forward-looking capabilities; (iv) Analysis of experimental results Experimental results show that: 1. In terms of corrosion type identification The recognition accuracy of the method of this invention is significantly higher than that of traditional empirical methods and single machine learning methods, and it is also superior to deep learning methods that do not introduce mechanistic constraints. 2. In terms of corrosion rate prediction The prediction error (RMSE) of the method of the present invention is significantly reduced, the prediction results are more stable, and they are more consistent with the actual corrosion trend. 3. Regarding the assessment of remaining useful life The remaining service life calculated by the method of this invention is closer to the actual maintenance cycle, and the error is significantly smaller than that of the comparative method. 4. In terms of early warning capabilities The method of this invention can detect corrosion anomalies in advance, significantly advancing the warning time and helping maintenance personnel to take timely measures. (V) Analysis of Mechanism Constraints Further analysis revealed: Deep learning methods that do not incorporate mechanistic constraints may exhibit significant prediction bias under certain operating conditions. This invention introduces a corrosion rate calculation formula. v(t) = (D prev - D curr ) / Δt Constrain the prediction results to make them conform to physical laws; This significantly improves the reliability and engineering applicability of the model.

[0136] (vi) Verification of closed-loop optimization effect In this embodiment, incremental model training is performed by introducing maintenance feedback data, and the performance changes before and after the model update are compared. The results show that: After the model was updated, the accuracy of corrosion type identification was further improved; The error in corrosion rate prediction has been further reduced; The model's adaptability to new operating conditions has been significantly enhanced; This demonstrates that the closed-loop optimization mechanism constructed in this invention can effectively improve the long-term performance of the system.

[0137] The comparative experiments described above demonstrate that the method of the present invention has the following significant advantages compared to existing technologies: Multi-source data fusion improves information utilization; Deep learning models have improved feature extraction and recognition capabilities; Mechanistic constraints enhance the physical consistency of prediction results; The closed-loop learning mechanism enhances the system's continuous optimization capabilities; This enables more accurate diagnosis and more reliable prediction and assessment of the corrosion status of atmospheric pressure towers.

[0138] 10. Example 9: Specific Application Example In this embodiment, the intelligent diagnosis method and system for atmospheric distillation tower corrosion faults based on the fusion of deep learning and mechanistic constraints described in this invention are applied to the corrosion monitoring and early warning scenario of an atmospheric distillation unit in an oil refinery.

[0139] (I) Application Background The atmospheric distillation tower of the refinery operates in a high-temperature, high-pressure and sulfur-containing environment for a long time. Key parts of the tower (such as the tower wall, trays and feed area) have varying degrees of corrosion problems. Traditional methods of relying on manual inspection and periodic testing are difficult to grasp the corrosion development trend in a timely manner, which poses certain safety risks.

[0140] (II) System Deployment In this embodiment, data acquisition units 100 are arranged at multiple key locations in the atmospheric pressure tower, including: Ultrasonic thickness sensor, used to collect wall thickness data; Temperature and pressure sensors are used to collect operating parameters; Vibration sensors are used to monitor the operating status of equipment; Near-infrared spectroscopy analyzer, used to detect the composition of a medium; Multiple measuring points were set up to cover corrosion-sensitive areas.

[0141] Furthermore, the data is transmitted to the edge processing unit 200 via an industrial communication network and then uploaded to the cloud analysis unit 300.

[0142] (III) Method Implementation Process In this embodiment, the system operates according to the following process: First, the wall thickness, temperature, pressure, vibration and medium composition data of each measuring point are acquired through the data acquisition unit 100, and a spatiotemporal data sequence with timestamps and spatial location identifiers is constructed. Then, in the edge processing unit 200, the data is processed by timestamp alignment, outlier removal, noise reduction and standardization, and multidimensional time series samples are constructed based on the sliding time window. Furthermore, the multi-source data is fused in the cloud analysis unit 300 to construct a time-series feature vector, and the corrosion evolution features are extracted through the CNN-LSTM model. The corrosion type is then identified using a deep neural network model. Based on this, a time series prediction model is used to predict the future corrosion rate, obtaining the predicted corrosion rate v_pred, and combining it with wall thickness data according to the formula: v(t) = (D prev - D curr ) / Δt Constrain or correct the prediction results; Subsequently, based on the current wall thickness value Dcurr, the critical wall thickness value D_threshold, and the predicted corrosion rate v_pred, according to the formula: RUL = (D curr - D threshold ) / v pred Calculate the remaining service life of each measuring point; Furthermore, based on the corrosion type identification results, corrosion rate, and remaining service life, the equipment operating status is classified and warned, and visualized through a digital twin model; Finally, after the equipment is repaired, the measured data is fed back to the model update unit 500 to perform incremental training on the model and achieve continuous optimization.

[0143] (iv) Application effect In this embodiment, by employing the method and system of the present invention, the following effects are achieved: It can continuously monitor the corrosion status of atmospheric pressure towers online; It can accurately identify different types of corrosion. It can dynamically predict corrosion rates and detect abnormal trends in advance; It can quantitatively assess the remaining service life, providing a basis for maintenance decisions; It can provide early warnings before corrosion risks occur, thus improving equipment operational safety; Furthermore, compared to traditional methods, this invention can detect potential corrosion risks in advance, thereby effectively reducing unplanned downtime.

[0144] (v) Engineering application value Verification through the above application examples shows that the method and system of the present invention are applicable to actual oil refining plant environments, have good engineering applicability and promotion value, and can significantly improve the level of intelligence in equipment corrosion monitoring and management.

[0145] It should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, various modifications, substitutions, or variations can be made to the above embodiments without departing from the principles and spirit of the present invention, and all such modifications, substitutions, or variations should fall within the scope of protection of the present invention.

[0146] Furthermore, the various technical features involved in this invention can be combined arbitrarily without conflict to form different implementation methods, all of which fall within the protection scope of this invention.

[0147] It should be understood that terms such as "comprising" or "including" used in this specification are intended to indicate the presence of the described features, steps, or components, but do not exclude the presence of other features, steps, or components; terms such as "preferred," "further," and "optional" are used only to describe different implementations and do not imply a limitation on the technical solution.

[0148] The scope of protection of this invention shall be determined by the claims. The description and drawings are only used for interpreting and explaining the claims.

Claims

1. A method for intelligent diagnosis of corrosion faults in atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints, characterized in that, Includes the following steps: S1. Collect wall thickness, temperature, pressure, vibration and medium composition data at multiple spatial locations in the atmospheric pressure tower, and construct a spatiotemporal data sequence containing timestamps and spatial location identifiers; S2. Perform timestamp alignment, outlier removal, noise reduction and standardization on the spatiotemporal data sequence, and aggregate the data according to a preset time window to form a multidimensional time series sample. S3. Based on timestamps and location identifiers, multi-source data are fused to construct a time-series feature vector containing operating parameters and corrosion characteristics; S4. Input the temporal feature vector into the combined model of convolutional neural network and long short-term memory network for feature extraction to obtain the erosion evolution feature representation; S5. Based on a deep neural network model, the corrosion evolution feature representation is used to identify the corrosion type, and based on a time series prediction model, the future corrosion rate is predicted. S6. Calculate the remaining service life based on the current wall thickness, predicted corrosion rate, and preset safety threshold, and use the corrosion rate mechanism formula to constrain or correct the prediction results; conduct early warning assessment based on corrosion type identification results, corrosion rate prediction results, and remaining service life, and incrementally train the deep learning model based on the measured data after maintenance to achieve adaptive model updates.

2. The method according to claim 1, characterized in that: The multidimensional time series samples are constructed using a sliding time window, which contains data from multiple consecutive time steps.

3. The method according to claim 1, characterized in that: The time-series feature vector includes wall thickness, temperature, pressure, vibration amplitude, vibration frequency, medium sulfur content, and flow velocity parameters.

4. The method according to claim 1, characterized in that: The corrosion rate is calculated according to the following formula: v(t) = (D prev - D curr ) / Δt, Where v(t) represents the corrosion rate at time t, D prev D represents the wall thickness value at the previous detection time. curr Δt represents the wall thickness value at the current detection time, and Δt represents the time interval between the previous detection time and the current detection time. And it serves as the basis for model training labels or prediction result correction.

5. The method according to claim 1, characterized in that: The remaining useful life is calculated according to the following formula: RUL = (D curr - Dt hreshold ) / v pred , Where RUL represents the remaining useful life, D curr This represents the wall thickness value at the current detection moment, Dthreshold represents the safety critical wall thickness value, and v pred This indicates the predicted corrosion rate; It is dynamically updated based on real-time monitoring data.

6. The method according to claim 1, characterized in that: It also includes comprehensive confidence calculation and data augmentation, wherein the comprehensive confidence is calculated as follows: Conf = α × Acc + β × (1 - RMSE / D curr ); Where Conf represents the overall confidence level, α represents the weighting coefficient of the corrosion type identification result, β represents the weighting coefficient of the corrosion rate prediction result, and α + β = 1, Acc represents the corrosion type identification accuracy, RMSE represents the root mean square error of the corrosion rate prediction, and D curr This indicates the wall thickness value at the current detection moment; The data augmentation involves using generative adversarial networks to generate synthetic fault samples to expand the training dataset.

7. A smart diagnostic system for corrosion faults in atmospheric pressure towers based on the fusion of deep learning and mechanistic constraints, characterized in that, include: Data acquisition unit, edge processing unit, cloud analysis unit, early warning and visualization unit, and model update unit; The data acquisition unit is used to collect wall thickness, temperature, pressure, vibration and medium composition data at multiple spatial locations in the atmospheric pressure tower, and to construct a spatiotemporal data sequence with timestamps and location identifiers. The edge processing unit is communicatively connected to the data acquisition unit and is used to perform timestamp alignment, outlier removal, noise reduction and standardization on the spatiotemporal data sequence, and generate multidimensional time series samples according to a preset time window. The cloud analysis unit is communicatively connected to the edge processing unit and is used to perform fusion modeling on the multidimensional time series samples, extract corrosion evolution features, perform corrosion type identification, corrosion rate prediction and remaining service life calculation, and constrain or correct the prediction results based on the corrosion mechanism formula. The early warning and visualization unit is communicatively connected to the cloud analysis unit and is used to trigger early warnings based on corrosion rate, remaining service life and diagnostic results, and to visualize the spatial location. The model update unit is communicatively connected to the cloud analysis unit and is used to receive maintenance data and perform incremental training and parameter optimization on the model.

8. The system according to claim 7, characterized in that: The data acquisition unit includes an ultrasonic thickness sensor, a temperature sensor, a pressure sensor, a vibration sensor, a near-infrared spectrometer, and an interface for reading process parameters from the distributed control system.

9. The system according to claim 7, characterized in that: The edge processing unit adopts an industrial edge computing gateway and is configured with local early warning rules to directly trigger on-site alarms without going through the cloud when the corrosion rate exceeds a preset threshold.

10. The system according to claim 7, characterized in that: The cloud-based analysis unit also includes a model interpretability module, which is used to calculate the contribution of input features to the diagnostic results based on Shapley values ​​or the Locally Interpretable Model Irrelevant Interpretation (LIME) algorithm.

Images