A multi-modal collaborative corrosion state monitoring method and system

Abstract

The present application relates to the technical field of material analysis, in particular to a multi-modal cooperative corrosion state monitoring method and system. Through a self-supervised contrast learning method of acoustic emission and electrochemical noise dual-mode signals, and combined with the innovative network architecture of Transformer and multilayer perception, the acoustic emission and electrochemical noise signal features are aligned, the problem of relying on a large amount of labeled data in the corrosion monitoring of chlor-alkali production equipment is overcome, and an online corrosion state monitoring model is formed. Under the self-supervised contrast learning framework, the model can use unlabeled acoustic emission and electrochemical noise data pairs for self-learning, reducing the demand for expensive and scarce labeled resources. Only a small amount of artificially labeled acoustic emission data is needed for fine-tuning training, and online corrosion state monitoring can be realized. At the same time, the present application can complete the on-site rapid and non-destructive corrosion state monitoring analysis of chlor-alkali production equipment only with acoustic emission data, promoting the development of corrosion state monitoring technology in the direction of more intelligent and automatic.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material analysis, and particularly relates to a multi-modal cooperative corrosion state monitoring method and system. BACKGROUND

[0002] In chlor-alkali chemical production, multiple dangerous chemicals such as hydrogen, chlorine, and chloroethylene are used and handled, which belong to the category of dangerous chemicals supervised by the state. Due to the complexity of the medium and the strong chemical activity in chlor-alkali chemical production, the production equipment is usually exposed to a corrosive environment of high chlorine (such as saturated brine, wet chlorine, and hypochlorite) and strong acid and alkali (such as hydrochloric acid and sodium hydroxide) for a long time. In order to cope with these harsh working conditions, 304, 316L stainless steel, TA2 titanium material, and Ni6 nickel material are usually used. However, even these materials with strong corrosion resistance are still susceptible to corrosion by chlorine ions and alkali solutions during long-term service, leading to pitting corrosion, stress corrosion, and other damage phenomena. Corrosion damage not only reduces the service life of the equipment, increases maintenance and replacement costs, but also may cause leakage, explosion, and other safety accidents, directly threatening the safety and stability of production. Therefore, corrosion has been one of the major challenges faced by the chlor-alkali industry. In order to ensure the safety and efficiency of chlor-alkali production, it is urgent to develop efficient and accurate monitoring technology for corrosion problems to monitor and warn the corrosion status of equipment in real time. Through effective corrosion monitoring technology, measures can be taken in time before the equipment is severely damaged, prolonging the service life of the equipment and reducing the safety risks caused by corrosion, thereby ensuring the safety and continuity of production.

[0003] Electrochemical noise technology is based on the real-time measurement of the spontaneous potential and current fluctuations generated during the corrosion process to determine the corrosion rate and corresponding mechanism. The data contains rich information, and through the analysis of the statistical characteristics of the noise signal, rich corrosion information can be obtained. However, the signal is usually complex, and its physical meaning is not easy to directly explain, which requires complex data processing and analysis methods, increasing the difficulty of online corrosion state monitoring, and making it difficult to arrange in the production site. Acoustic emission technology is based on the detection of elastic waves generated during the mechanical damage process of the specimen surface. The advantage is that the acoustic transducer is placed far away from the corrosion location, which can realize non-destructive testing and is convenient for production site arrangement. The disadvantage is that the data interpretation of the sound signal is complex, and there may be a delay between the acoustic emission event and the expected corrosion process.

[0004] A single monitoring method can only reflect some aspect of the corrosion process, and cannot comprehensively and accurately evaluate the corrosion state. Combining electrochemical noise and acoustic emission signals can simultaneously obtain chemical reaction and physical change information, improving the overall understanding of the corrosion process. However, acoustic emission and electrochemical signals have different time-frequency characteristics, and there is currently no method for accurately identifying and monitoring corrosion by combining acoustic emission and electrochemical noise. There is also a lack of correlation and mutual verification relationship between acoustic emission and electrochemical noise signals. Traditional corrosion analysis methods rely on a large amount of labeled data for supervised learning, which is not only time-consuming but also costly. The acquisition of labeled data is often limited by the experience and time of experts, which limits the ability to quickly and accurately analyze corrosion.

[0005] To solve the above problems, a multi-modal collaborative corrosion state monitoring method and system are provided. This method represents the feature space of electrochemical noise and acoustic emission signals through multi-level alignment, and comprehensively analyzes various characteristic information generated during the corrosion process, which can more comprehensively and accurately reflect the corrosion state. A self-supervised pre-training algorithm is used to allow the model to learn the intrinsic characteristics of the corrosion signal through self-learning, reducing the dependence on a large amount of labeled data. Through the model learning process, electrochemical noise information is stored in the model, and only acoustic emission devices need to be set up to collect signals during on-site monitoring, thereby achieving on-site non-destructive monitoring. SUMMARY

[0006] To solve the problems in the above background art, the present application provides a multi-modal collaborative corrosion state monitoring method and system based on electrochemical noise signals and acoustic emission signals to achieve rapid, non-destructive, accurate and automated monitoring of corrosion damage in chlor-alkali equipment.

[0007] To achieve the above-mentioned purpose, the scheme provided by the present application is as follows:

[0008] S1: Signal acquisition: The sample material includes but is not limited to stainless steel and carbon steel of various materials, which is placed in a corrosion solution to undergo corrosion. The solution used includes but is not limited to NaCl solution and other corrosion solutions. An electrochemical acquisition device and an acoustic emission acquisition device are set up on the test piece to simultaneously collect electrochemical noise signals and acoustic emission signals.

[0009] S2: Data set construction: Using a large amount of collected electrochemical noise signals and acoustic emission signals, first align the two signals at the time step to construct a large-scale acoustic-electric alignment pre-training data set without manual labeling; a small amount of acoustic emission signals with obvious corrosion characteristics are manually labeled to construct a labeled data set for the corrosion stage of acoustic emission.

[0010] S3: Model construction and training: an improved two-stage network is constructed, the first stage is a contrast learning structure based on Transformer, and the second stage is a corrosion stage monitoring network composed of a Transformer-based feature extractor and a multilayer perceptron-based classifier. The two-stage network is trained in turn using the two data sets constructed in step S2.

[0011] S4: Model testing: the improved corrosion stage monitoring model is used to judge the unknown acoustic emission test signal, and the corrosion stage corresponding to the test signal is output, to obtain the test effect of the corrosion monitoring model.

[0012] S5: System building: an interactive interface is made, software is packaged, and the software is installed in a server; the on-site acoustic emission signal acquisition device is connected to the packaged software to realize the on-site corrosion state monitoring system.

[0013] As a further scheme of the application: the material is a commonly used material in chlor-alkali production equipment, such as 304 and 316L stainless steel and carbon steel.

[0014] As a further scheme of the application: in step S1, the commonly used materials 304 and 316L stainless steel and carbon steel in chlor-alkali production equipment are used, and a large amount of rich data of different materials are used to complete the training, to help the model have better generalization ability.

[0015] As a further scheme of the application: the data set construction in step S2 includes the following steps:

[0016] S21: Constructing acoustic-electric alignment pre-training data set: without manual corrosion stage judgment, the artificial analysis cost is greatly saved, first, the collected electrochemical noise signal and acoustic emission signal are aligned in time steps, and are segmented into small fragments with suitable time length, to form a large number of acoustic-electric alignment data pairs, which are used as the acoustic-electric alignment pre-training data set.

[0017] S22: Constructing acoustic emission corrosion stage labeled data set: for acoustic emission data, a small amount of acoustic emission signals with obvious corrosion characteristics are manually labeled to construct the acoustic emission corrosion stage labeled data set. In addition, in addition to the above data set, part of the acoustic emission data set is additionally collected to form a test data set, which is used for subsequent model effect evaluation.

[0018] As a further scheme of the application: the model construction in step S3 includes the following steps:

[0019] S31: Constructing a Transformer-based feature extractor: Utilize self-attention mechanisms to enable the model to focus on the temporal information in the electrochemical noise and acoustic emission time series.

[0020] S32: Constructing a model learning strategy based on the first-stage self-supervised contrast learning architecture: Train the feature extractor proposed in S31, and use the acoustic-electric data pairs in the pre-training dataset for contrast learning pre-training to train the model to achieve collaborative learning of acoustic emission signals and electrochemical noise signals, and store the effective information of the electrochemical noise signals, with a distribution alignment loss and an event alignment loss as the loss function.

[0021] S33: Constructing a model learning strategy for the second-stage corrosion stage monitoring: Extract the acoustic emission signal feature extractor trained in step S32, and combine a classifier based on a multilayer perception mechanism to further supervise the training based on the labeled data set of the acoustic emission corrosion stage in step S2, to train the model to obtain effective monitoring corrosion stage capability.

[0022] As a further scheme of the present application: In step S4, the model test uses the acoustic emission data of the test data set as input, and based on the model output result, determines the corrosion state of each sample in the test set, and calculates the model detection accuracy for evaluation.

[0023] As a further scheme of the present application: In step S5, the system is built by deploying the above corrosion monitoring model to a server and connecting to an acoustic emission signal acquisition module, and after the signal acquisition module acquires the signal, the data is preprocessed in the signal preprocessing module and transmitted to the corrosion state monitoring algorithm in the server, and the online corrosion state monitoring result is calculated.

[0024] The beneficial effects of the present application are:

[0025] 1. Compared with the supervised deep learning method relying on a large amount of labeled data, the self-supervised learning algorithm is used, a large amount of acoustic-electric data pairs without manual annotation are used to pre-train the model, and the dependence on labeled data is greatly reduced.

[0026] 2. By designing various alignment objective functions in the alignment learning stage of acoustic emission and electrochemical noise signals, the model realizes effective alignment of acoustic emission and electrochemical noise signals, fully utilizes the related information of the two signals, and improves the corrosion monitoring capability of the model.

[0027] 3. The present application can perform on-site, non-destructive and efficient corrosion state monitoring on chlor-alkali production equipment. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1A network structure diagram of a multi-modal cooperative corrosion state monitoring method and system of the present application.

[0029] Figure 2 A flowchart of a multi-modal cooperative corrosion state monitoring method and system of the present application. DETAILED DESCRIPTION

[0030] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings. The following examples or drawings are used to illustrate the present application, but not to limit the scope of the present application.

[0031] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings. The following examples or drawings are used to illustrate the present application, but not to limit the scope of the present application.

[0032] REFERENCE Figure 2 A multi-modal cooperative corrosion state monitoring method and system, including specific methods comprising steps S1-S5:

[0033] S1: Signal acquisition: collect signals of two modalities of electrochemical noise and acoustic emission.

[0034] For example, the experimental temperature is set to room temperature; the solution uses a corrosion solution such as NaCl solution; the sample material uses stainless steel and carbon steel of various materials. The electrochemical noise acquisition method refers to ISO 17093-2015 Corrosion of metals and alloys-Guidelines for corrosion test by electrochemical noise measurements, and the experimental equipment uses Gamry Reference600; the acoustic emission acquisition device uses a PCI-2 acoustic emission detection system of the American Physical Acoustics Company (PAC) with a gain of 40 dB, and a WD type acoustic emission broadband sensor is arranged above the test piece.

[0035] S2: Data set construction: first, the collected electrochemical noise signals and acoustic emission signals are aligned in time steps, and are segmented into small segments of suitable time length to form a large number of acoustic-electric alignment data pairs for use as acoustic-electric alignment pre-training data sets. For the acoustic emission data, a small amount of acoustic emission signals with obvious corrosion characteristics are manually labeled to construct a labeled data set of the acoustic emission corrosion stage. The sample corrosion state samples mainly include no corrosion, metastable corrosion stage and stable corrosion stage. In addition, in addition to the above data sets, part of the acoustic emission data sets are additionally collected to constitute a test data set for subsequent evaluation of the effect of the model. To ensure the test effect, the training set and the test set samples should not overlap.

[0036] Exemplary, first, the electrochemical noise and acoustic emission signals collected in step S1 are aligned in time steps, segmented into small pieces of 10 min time series data, and a large number of acoustic and electric data pairs are constructed. Among them, a small amount of acoustic emission signals with obvious corrosion characteristics are manually labeled to construct a labeled acoustic emission corrosion stage data set. The sample corrosion state sample mainly includes non-corrosion, metastable corrosion stage and stable corrosion stage. In addition, in addition to the above data set, part of the acoustic emission data set is additionally collected to form a test data set, which is used for subsequent model effect evaluation.

[0037] S3: Model construction and training: Referring to Figure 1 , an improved network is constructed. First, a time series self-attention feature extractor is constructed based on Transformer, and a two-stage learning strategy is designed. The first stage constructs a model learning strategy based on a self-supervised contrast learning architecture, and the second stage constructs a model learning strategy for supervised training of the corrosion stage.

[0038] The training data set constructed in step S2 is used for two-stage network training.

[0039] S31: A time series self-attention feature extractor is constructed based on Transformer, which performs attention calculation between time series samples in the time dimension direction and extracts time series information.

[0040] S32: Construct a first-stage model learning strategy based on a self-supervised contrast learning architecture: as shown in the accompanying Figure 1 , the self-supervised contrast structure has multiple parallel branches for two modal data, each branch includes the feature extractor described in S31, and is connected to a multi-layer perceptron structure after that. The feature representation vectors output by each branch are respectively calculated by each collaborative alignment loss function to optimize the model parameters. Sample event alignment calculation, q evt represents the extracted acoustic emission signal feature, k evt represents the extracted acoustic emission signal feature, y represents h EN extracted electrochemical noise signal feature, and its formula is L evt = L nce (ζ(q evt ,y),{ζ(k evt ,y)}) where Two modal distribution alignment calculation, its formula is where

[0041] S33: Construct a second-stage model learning strategy for supervised corrosion stage monitoring: extract the acoustic emission feature extractor trained in the first stage A classifier based on multi-layer perception is added thereafter. The model is input with the labeled data set samples of acoustic emission corrosion stage, and the classification cross-entropy loss is calculated to train the model to obtain the corrosion stage classification ability, and the formula is

[0042] S4: Corrosion monitoring capability evaluation: the test data set is classified by the trained model to obtain the evaluation result of the model detection performance.

[0043] For example, according to the acoustic emission data of the test set, the prediction result of the model is calculated The accuracy between the true label y is used for the evaluation of the prediction ability of the model, and the formula is Where Accuracy represents the accuracy, TP represents the true positive, TN represents the true negative, FP represents the false positive, and FN represents the false negative.

[0044] S5: System building: making an interactive interface, packaging software, installing software in a server; connecting the on-site acoustic emission signal acquisition device with the packaged software to realize the on-site corrosion state monitoring system.

[0045] The system includes a signal acquisition device, a processor and a memory, wherein the memory stores computer program instructions, and when the processor executes the computer program instructions, the corrosion state monitoring method based on acoustic emission signals is realized. The system acquires on-site acoustic emission signals through the signal acquisition device, and transmits the signals to the processor through the communication interface. The processor pre-processes and extracts features from the signals, and then analyzes the corrosion state of the equipment. Other components of the system, such as communication protocols and data storage methods, are well known to those skilled in the art, so they will not be described here.

[0046] Compared with the supervised deep learning method relying on a large amount of labeled data, the self-supervised learning algorithm is adopted, and a large amount of acoustic emission and electrochemical noise training data without manual annotation is used for model pre-training, which significantly reduces the demand of the model for labeled data. The feature extractor based on Transformer combined with the classifier based on multi-layer perception fully learns the effective time sequence information of the time sequence data, effectively improves the performance of the image segmentation task. At the same time, this method can complete the corrosion state monitoring task only by using acoustic emission data, and is especially suitable for on-site non-destructive corrosion state monitoring of chlor-alkali production equipment.

[0047] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the same; although the present application has been described in detail with reference to the foregoing examples, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing examples can be modified, or some technical features thereof can be replaced by equivalent ones; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A multimodal collaborative corrosion state monitoring method, characterized in that: The specific method includes the following steps: S1: Signal Acquisition: The sample material includes stainless steel and carbon steel, which are placed in a corrosion solution to undergo corrosion. The corrosion solution used includes NaCl solution. An electrochemical acquisition device and an acoustic emission acquisition device are set on the sample to simultaneously acquire electrochemical noise signals and acoustic emission signals. S2: Dataset Construction: Using the collected electrochemical noise signal and acoustic emission signal, the two signals are first aligned at the time step to construct a large-scale acoustic-electrochemical alignment pre-training dataset that does not require manual annotation; acoustic emission signals with obvious corrosion characteristics are manually annotated to construct a labeled dataset for the acoustic emission corrosion stage. S3: Model Building and Training: Construct an improved two-stage network. The first stage is a contrastive learning structure based on Transformer, and the second stage is an erosion stage monitoring network composed of a feature extractor based on Transformer and a classifier based on multilayer perceptron. Train the two-stage networks sequentially using the two datasets constructed in step S2. S4: Model Testing: The improved corrosion stage monitoring model is used to determine the unknown acoustic emission test signal and output the corrosion stage corresponding to the test signal. S5: System Setup: Create the interactive interface, package the software, and install the software on the server; Connect the on-site acoustic emission signal acquisition device with the packaging software to realize the on-site corrosion status monitoring system; The model construction in step S3 includes the following steps: S31: Construct a Transformer-based feature extractor: Utilize the self-attention mechanism to enable the model to focus on temporal information in electrochemical noise and acoustic emission time series; S32: Constructing the first-stage model learning strategy based on a self-supervised contrastive learning architecture: train the feature extractor proposed in S31, use the acoustic and electrical data pairs in the pre-training dataset for contrastive learning pre-training, train the model to achieve collaborative learning of two modal signals, acoustic emission signal and electrochemical noise signal, store the effective information of electrochemical noise signal, and its loss function is distribution alignment loss and event alignment loss. S33: Model learning strategy for the second stage of corrosion monitoring: Extract the acoustic emission signal feature extractor trained in step S32, and combine it with a classifier built on a multilayer perceptron. Further supervised training is performed on the labeled dataset of acoustic emission corrosion stage in step S2 to train the model to obtain the ability to effectively monitor the corrosion stage.

2. The multimodal collaborative corrosion state monitoring method according to claim 1, characterized in that: In step S1, signal acquisition used 304 and 316L stainless steel and carbon steel, which are commonly used in chlor-alkali production equipment. A large amount of rich data from various materials was used to complete the training, which helped the model have better generalization ability.

3. The multimodal collaborative corrosion state monitoring method according to claim 1, characterized in that: The dataset construction in step S2 includes the following steps: S21: Constructing an acoustic-electric alignment pre-training dataset: Eliminating the need for manual corrosion stage judgment saves on manual analysis costs. First, the collected electrochemical noise signal and acoustic emission signal are aligned in time step and divided into smaller segments of suitable time length to form a large number of acoustic-electric alignment data pairs, which are used as an acoustic-electric alignment pre-training dataset. S22: Construct a labeled dataset for the acoustic emission corrosion stage: For acoustic emission data, acoustic emission signals with obvious corrosion characteristics are manually labeled to construct a labeled dataset for the acoustic emission corrosion stage; In addition, besides the above dataset, additional acoustic emission datasets are collected to form a test dataset for subsequent evaluation of model performance.

4. The multimodal collaborative corrosion state monitoring method according to claim 1, characterized in that: In step S4, the model test takes the acoustic emission data of the test dataset as input, and based on the model output results, determines the corrosion state of each sample in the test set, and calculates the model detection accuracy for evaluation.

5. The multimodal collaborative corrosion state monitoring method according to claim 1, characterized in that: The system setup in step S5 involves deploying the corrosion monitoring model trained in step S3 to the server and connecting it to the acoustic emission signal acquisition module. After the acoustic emission signal acquisition module acquires the signal, the data is preprocessed by the signal preprocessing module and then transmitted to the corrosion state monitoring algorithm on the server to calculate the online corrosion state monitoring results.

6. The multimodal collaborative corrosion state monitoring method according to claim 1, characterized in that: The sample material includes at least one of 304 stainless steel, 316L stainless steel and carbon steel, and the corrosion solution includes NaCl solution.

Images