Thermal power plant equipment fault prediction method and system based on multi-modal data

Abstract

The invention discloses a thermal power plant equipment fault prediction method based on multi-modal data, and relates to the technical field of thermal power plant fault prediction, and the prediction steps are as follows: S1, sound waves are emitted in a thermal power plant circulating water system based on a sound wave emitter, emitted elastic waves are received based on a sensor, the elastic waves are converted into electric signals, and the electric signals are preprocessed and then stored; s2, obtaining text information based on a historical maintenance report in the thermal power plant, screening maintenance information related to the circulating water system from the text information based on a regular expression, and obtaining a text feature vector in the maintenance information based on a TF-IDF algorithm; the invention further discloses a thermal power plant equipment fault prediction system based on the multi-modal data. According to the method, the real-time operation state and historical fault experience are integrated through multi-modal information fusion, the limitation of a single modal is made up, workers are assisted to find faults in advance, the probability of sudden faults is reduced, and the accuracy and reliability of prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal power plant fault prediction, and in particular to a thermal power plant equipment fault prediction method and system based on multimodal data. Background Art

[0002] Thermal power plants use combustible materials such as coal, oil, and natural gas as fuel to produce electricity. The chemical energy of the fuel is converted into heat energy through combustion, which is then converted into mechanical energy through a steam turbine. Finally, the mechanical energy is converted into electricity through a generator. The circulating water system plays a vital role in thermal power plants. It is mainly used to cool the steam in the condenser and condense it into water, thereby maintaining the vacuum degree of the condenser and improving the thermal efficiency of the unit. In addition, the circulating water system can also provide cooling water for other equipment that requires cooling, such as oil coolers and generator air coolers. However, the circulating water system is in a complex operating environment for a long time and is affected by multiple factors such as water quality, temperature, and pressure. The equipment is prone to failures such as pipeline corrosion and scaling, pump wear, and cavitation. Once the circulating water system fails, it will not only cause the unit to operate at a reduced load, but may even cause a shutdown accident, resulting in huge economic losses to the thermal power plant.

[0003] In the past, equipment failure prediction in thermal power plants was mostly limited to the use of single-modal data, among which the analysis of equipment operation electrical signals was the most common. Through time domain analysis, the amplitude, frequency, and duration characteristics of the electrical signals can be obtained, thereby making a preliminary judgment on the equipment operation status. However, this method has significant disadvantages: on the one hand, the electrical signal only reflects the electrical characteristics of the equipment operation and cannot touch upon the multi-dimensional factors behind the equipment failure, such as the degree of equipment aging and historical maintenance. For example, the historical maintenance report of the equipment records in detail the type, cause and maintenance plan of the past failure. This valuable information is of great significance for current failure prediction, but traditional electrical signal analysis methods find it difficult to extract value from it; on the other hand, the failure of the circulating water system of a thermal power plant is the result of the interweaving of multiple factors such as water quality, temperature, pressure, and mechanical properties of the equipment. A single electrical signal modality cannot fully capture these complex causal relationships, resulting in a significant reduction in the accuracy and reliability of fault prediction. To this end, we propose a thermal power plant equipment failure prediction method and system based on multi-modal data. Summary of the Invention

[0004] In order to solve the above technical problems, a method and system for predicting equipment faults in thermal power plants based on multimodal data are provided. This technical solution solves the problem of reduced accuracy in fault prediction.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting equipment failure in a thermal power plant based on multimodal data, wherein the prediction steps are as follows:

[0006] The prediction method for thermal power plant equipment failure based on multimodal data includes the following steps:

[0007] S1. An acoustic wave transmitter emits acoustic waves in the circulating water system of a thermal power plant, and a sensor receives the emitted elastic waves, converts them into electrical signals, and stores the electrical signals after preprocessing.

[0008] S2. Based on historical maintenance reports from the thermal power plant, obtain text information, filter out maintenance information related to the circulating water system from the text information using regular expressions, and obtain text feature vectors from the maintenance information using the TF-IDF algorithm;

[0009] S3. Analyze the electrical signal based on the time domain analysis method to obtain the law of signal change over time, and calculate the signal amplitude and signal duration to obtain the electrical signal characteristics;

[0010] S4, fusing the electrical signal features with the text feature vector based on the splicing method to form a multimodal comprehensive feature;

[0011] S5. Build a prediction model and pre-train the correlation between learning features and faults based on the cross-validation method; input multimodal comprehensive features into the prediction model, predict faults in the circulating water system of the thermal power plant, and output the prediction results;

[0012] The time domain analysis method in step S3 converts the electrical signal into a discrete time series and calculates and analyzes it through the autocorrelation function to obtain the law of signal change over time; the signal amplitude is obtained by calculating the peak value of the signal, and the signal duration calculation reflects the duration of the abnormal state of the equipment.

[0013] Preferably, in step S2, text information is obtained based on the equipment operation records and maintenance report text materials accumulated by the thermal power plant, and the text information is filtered using regular expressions. The filtering steps are: matching by scanning the text information line by line, extracting the context of the matching line, and obtaining maintenance information related to the circulating water system.

[0014] Preferably, the regular expression construction includes: a device name matching mode, a maintenance operation matching mode and a fault phenomenon matching mode; wherein the device name matching mode uses the device name of the circulating water system as a keyword to construct an expression; the maintenance operation matching mode uses the maintenance action as a keyword to construct an expression; and the fault phenomenon matching mode uses the fault phenomenon description as a keyword to construct an expression;

[0015] Line-by-line scanning matches each piece of text information obtained, splits the text information into lines, matches it line by line with the constructed regular expression, and uses the successfully matched text information as maintenance information related to the circulating water system.

[0016] Preferably, the steps based on the TF-IDF algorithm in step S2 are:

[0017] TF-IDF calculation: Use natural language processing tools to segment the selected maintenance information, split the text into individual words, count the number of times each word appears in the text, obtain the original word frequency, and normalize the word frequency using the following formula:

[0018]

[0019] Among them TF ij Indicates the frequency of the jth word in the i-th text, n ij is the number of occurrences of the jth word in the i-th text, is the sum of the number of times all words appear in the i-th text;

[0020] Determine the inverse document frequency: traverse the entire maintenance information text collection, count the number of documents containing each word, and calculate the inverse document frequency;

[0021] The collection of words related to thermal power plant equipment is V e , for each vocabulary v∈V in the set e , assign weighting factor α v , where α v ≥1, for non-thermal power plant equipment related words, the weighting factor α v =1;

[0022] Introducing the thermal power plant equipment dictionary D for the keyword w k , w k ∈D, adjust the document frequency statistics and introduce a correction factor β, then the corrected document frequency is

[0023] For non-keyword w u ,Right now The document frequency remains constant, i.e.

[0024] For the keyword w k The inverse document frequency calculation formula is:

[0025]

[0026] For non-keyword w n The inverse document frequency calculation formula is:

[0027]

[0028] Combined with the weighting factor α of the vocabulary related to thermal power plant equipment w , the adjusted TF-IDF calculation formula is:

[0029] TF-IDF w,d =α w ×TF w,d ×IDF w

[0030] in For keyword w k The inverse document frequency measures the importance of a word in a document collection. σ is the total number of documents in the text collection, that is, the number of documents involved in the calculation. is the number of documents originally containing the keyword wk, is the number of documents containing the keyword wk after correction; is the number of documents without keyword wu, No document frequency correction is done for non-keywords, i.e. equal

[0031] Preferably, the time domain analysis method step in step S3 is:

[0032] The electrical signal is recorded as a discrete time series x[n'], n'=0,1,2,…,N'-1, where N' is the number of signal sampling points and the sampling interval is T s ;

[0033] The signal periodicity is analyzed by the autocorrelation function, and the calculation expression is:

[0034]

[0035] where R xx [m] is the autocorrelation function of the signal's temporal variation, where m is the number of delay points and the value range of m is -(N'-1)≤m≤(N'-1);

[0036] Signal amplitude calculation: The signal amplitude is determined by calculating the peak value of the signal. In a discrete signal, the maximum value x of the signal is determined by traversing all sampling points. max and the minimum value x min , the signal amplitude calculation formula is:

[0037] A=max(|x max |,|x min |)

[0038] Where A is the calculated signal amplitude.

[0039] Preferably, the signal duration calculation in step S3 is: suppose the signal starts from the starting time From the beginning to the end End, the signal amplitude is greater than the threshold x during this time period th , by traversing the signal sequence, find The first point and the last point The calculation formula based on the duration B of continuous signal interruption is:

[0040]

[0041] Where B is the calculated signal duration.

[0042] Preferably, the splicing method in step S4 is as follows: perform dimensionality check on the feature vectors extracted from the maintenance report text and the feature vectors extracted from the electrical signal, adjust the vector dimensions of the two groups based on the zero-padding method to keep the number consistent; perform a concat operation on the electrical signal features and the text feature vectors on the feature dimension to obtain a comprehensive feature vector of multimodal information.

[0043] Preferably, the cross-validation process in step S5 includes:

[0044] (a) Divide the sample set into a training set and a validation set using the k-fold cross-validation method;

[0045] (b) Construct a prediction model based on a deep neural network. The network structure includes an input layer, three fully connected hidden layers, and an output layer, and the activation function is ReLU;

[0046] (c) Using the cross entropy loss function as the optimization objective, combined with the Adam optimization algorithm to dynamically adjust the learning rate, and update the network weight parameters through back propagation;

[0047] (d) Repeat k rounds of cross-validation, with k set to 5 and 10, and perform k rounds of cycles. In each round, 1 fold is selected as the validation set, and the remaining k-1 folds are combined into the training set.

[0048] (e) Calculate the average F1-score of each round and select the model parameters with the highest average F1-score as the final model;

[0049] (f) Use the test set to evaluate the model and calculate the evaluation indicators on the test set. If there are differences, the model is readjusted. If there are no differences, the trained prediction model is obtained.

[0050] Preferably, in step S5, after determining that there is a fault risk, the model gives a predicted fault type to ensure stable operation of the circulating water system of the thermal power plant; the predicted fault types include pipeline failure, pump wear, pump cavitation and fan failure.

[0051] The thermal power plant equipment failure prediction system based on multimodal data includes:

[0052] an electrical signal collection module configured to collect sound waves in the circulating water system and convert them into electrical signals;

[0053] A historical maintenance information collection module configured to collect and process historical maintenance reports in the thermal power plant;

[0054] an electrical signal characteristic analysis module, configured to perform characteristic analysis on the electrical signal;

[0055] A fusion module is configured to fuse the extracted electrical signal features with the feature vectors extracted from the maintenance report text based on a splicing method to form a comprehensive feature of multimodal information;

[0056] The prediction module is configured to perform pre-training learning based on the correlation between learning features and faults using the cross-validation method, input the comprehensive features of multimodal information into the model for prediction output, and predict faults in the circulating water system of the thermal power plant.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The present invention acquires elastic wave electrical signals through acoustic wave transmitters and sensors, processes historical maintenance reports with regular expressions and the TF-IDF algorithm, mines out past equipment failure frequencies, failure points, and empirical knowledge of maintenance measures, and combines real-time electrical signals to analyze the entire life cycle of the equipment. Multimodal information fusion integrates real-time operating status and historical failure experience, making up for the limitations of a single modality, helping staff detect faults in advance, reducing the probability of sudden failures, and improving the accuracy and reliability of predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flow chart of the steps for predicting equipment failure in a thermal power plant according to the present invention;

[0060] Figure 2 This is a framework diagram of the thermal power plant equipment fault prediction system of the present invention. DETAILED DESCRIPTION

[0061] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0062] Reference Figure 1 As shown in the figure, the thermal power plant equipment failure prediction method based on multimodal data has the following prediction steps:

[0063] S1. An acoustic wave transmitter emits acoustic waves in the circulating water system of a thermal power plant, and a sensor receives the emitted elastic waves, converts them into electrical signals, and stores the electrical signals after preprocessing.

[0064] S2. Based on historical maintenance reports from the thermal power plant, obtain text information, filter out maintenance information related to the circulating water system from the text information using regular expressions, and obtain text feature vectors from the maintenance information using the TF-IDF algorithm;

[0065] S3. Analyze the electrical signal based on the time domain analysis method to obtain the law of signal change over time, and calculate the signal amplitude and signal duration to obtain the electrical signal characteristics;

[0066] S4, fusing the electrical signal features with the text feature vector based on the splicing method to form a multimodal comprehensive feature;

[0067] S5. Build a prediction model and pre-train the correlation between learning features and faults based on the cross-validation method; input multimodal comprehensive features into the prediction model, predict faults in the circulating water system of the thermal power plant, and output the prediction results;

[0068] The time domain analysis method in step S3 converts the electrical signal into a discrete time series and calculates and analyzes it through the autocorrelation function to obtain the law of signal change over time; the signal amplitude is obtained by calculating the peak value of the signal, and the signal duration calculation reflects the duration of the abnormal state of the equipment.

[0069] The acoustic wave transmitter and sensor in this application work together to collect elastic waves and convert them into electrical signals. This can sensitively capture subtle changes in the equipment's mechanical structure, supplement the mechanical data dimension, and provide first-hand information on the equipment's physical state for fault diagnosis. Regular expressions and the TF-IDF algorithm are used to process historical maintenance reports, uncovering past fault frequencies, fault points, and repair measures. These are then combined with real-time electrical signals, infusing long-term, valuable experience into fault prediction from the perspective of the equipment's entire lifecycle. Time-domain analysis is used to analyze electrical signals, visually presenting equipment operating dynamics through signal variation patterns, amplitude, and duration. Key fault indicator features are quickly extracted, facilitating the timely detection of electrical anomalies. Electrical signal features are then combined with maintenance report feature vectors to fully integrate real-time operating status with historical fault experience. The complementary nature of different modal data makes fault diagnosis more comprehensive and accurate. A predictive model is constructed and cross-validated to prevent overfitting. Leveraging comprehensive and accurate multimodal features, the model deeply learns feature associations with faults, outputting highly reliable prediction results. This helps personnel identify potential faults in advance, rationally plan maintenance, reduce the probability of sudden failures, ensure stable operation of the thermal power plant's circulating water system, and minimize downtime losses.

[0070] The data preprocessing in step S1 includes denoising, filtering and normalization.

[0071] This application adopts a wavelet threshold denoising algorithm. Based on the frequency band range of the acoustic signal of the circulating water system of a thermal power plant, its main frequency band range is between 10Hz and 500Hz. On this basis, the number of decomposition layers is set to 5, and the threshold function is selected as Sqtwolog. Through this algorithm, the noise interference in the acoustic signal is removed, making the signal smoother and truly reflecting the system state. The principle of wavelet threshold denoising is to perform multi-layer wavelet decomposition on the signal, and then process the decomposed coefficients according to the threshold, retaining the effective signal components and removing the coefficients corresponding to the noise. Finally, wavelet reconstruction is performed to obtain the denoised signal;

[0072] The Butterworth bandpass filter is used for filtering. Based on the characteristics of the acoustic wave signal of the thermal power plant's circulating water system, its effective frequency range is between 20Hz and 400Hz. Therefore, the passband range of the Butterworth bandpass filter is set to 20Hz to 400Hz. This can retain signal components with frequencies within the effective range while effectively filtering out high-frequency noise and power frequency interference. The Butterworth filter has a flat passband characteristic and can maintain the original form of the signal as much as possible within the passband, reducing signal distortion.

[0073] Normalization processing linearly maps the amplitude of the electrical signal to the range of [-1, 1]. Since the measurement ranges of different sensors may vary, this difference will affect subsequent data analysis and model training. Through normalization processing, the impact of sensor range differences can be eliminated, so that all signal amplitudes are in the same scale range, improving data comparability and model training effects.

[0074] In step S2, text information is obtained based on the equipment operation records and maintenance report text data accumulated by the thermal power plant, and regular expressions are used to filter the text information. The filtering steps are: matching by scanning the text information line by line, extracting the context of the matching line, and obtaining maintenance information related to the circulating water system.

[0075] Acquiring text information based on equipment operation records and maintenance report text materials can fully tap into the large amount of valuable data contained therein on the historical operation status of the equipment and the occurrence of faults, providing a rich reference basis for fault prediction; using regular expressions to scan text matching patterns line by line can accurately extract information related to the circulating water system, avoiding omissions and errors that may occur in manual screening, and improving the efficiency and accuracy of information extraction.

[0076] The principle of electrical signal conversion in step S1 is as follows: the sensor relies on the piezoelectric effect to convert elastic waves into electrical signals. The core of the sensor uses quartz crystals and piezoelectric ceramics. When it is mechanically deformed by external force, surface charges at both ends accumulate, and the amount of charge is proportional to the external force. The vibration energy carried by the elastic wave acts on the piezoelectric material, causing mechanical deformation. The frequency, amplitude and propagation direction characteristics of the elastic wave cause the piezoelectric material to produce compression, tension or shear deformation, breaking the internal charge balance. As the piezoelectric material deforms, the internal charge center displaces, and the surface charge separates and polarizes, generating a voltage difference that changes with the elastic wave vibration, forming a corresponding electrical signal. The signal conditioning circuit in the sensor amplifies and filters the weak electrical signal to improve signal quality and facilitate transmission analysis.

[0077] Regular expression construction includes: equipment name matching mode, maintenance operation matching mode and fault phenomenon matching mode. The equipment name matching mode uses the equipment name of the circulating water system as the keyword to construct the expression; the maintenance operation matching mode uses the maintenance action as the keyword to construct the expression; the fault phenomenon matching mode uses the fault phenomenon description as the keyword to construct the expression;

[0078] Line-by-line scanning matches each piece of text information obtained, splits the text information into lines, matches it line by line with the constructed regular expression, and uses the successfully matched text information as maintenance information related to the circulating water system.

[0079] This application is based on the processing of circulating water system related information of thermal power plant text data. Line-by-line scanning is to split each text data into lines, and match them line by line with three types of regular expression patterns. Any text that matches any type will be filtered out; extracting matching lines and context is centered on the matching text line, extracting the text content of the previous and next lines to form fragments, and doing this for all matching lines. The fragments are collected and classified to form a data set. These operations can accurately filter information, efficiently process text, and enrich information extraction, providing valuable data support for fault prediction.

[0080] The steps based on the TF-IDF algorithm in step S2 are:

[0081] TF-IDF calculation: Use natural language processing tools to segment the selected maintenance information, split the text into individual words, count the number of times each word appears in the text, obtain the original word frequency, and normalize the word frequency using the following formula:

[0082]

[0083] Among them TF ij Indicates the frequency of the jth word in the i-th text, n ij is the number of occurrences of the jth word in the i-th text, is the sum of the number of times all words appear in the i-th text;

[0084] Determine the inverse document frequency: traverse the entire maintenance information text collection, count the number of documents containing each word, and calculate the inverse document frequency;

[0085] The collection of words related to thermal power plant equipment is V e , for each vocabulary v∈V in the set e , assign weighting factor α v , where α v ≥1, for non-thermal power plant equipment related words, the weighting factor α v =1;

[0086] Introducing the thermal power plant equipment dictionary D for the keyword w k , w k ∈D, adjust the document frequency statistics and introduce a correction factor β, then the corrected document frequency is

[0087] For non-keyword w u ,Right now The document frequency remains constant, i.e.

[0088] For the keyword w k The inverse document frequency calculation formula is:

[0089]

[0090] For non-keyword w n The inverse document frequency calculation formula is:

[0091]

[0092] Combined with the weighting factor α of the vocabulary related to thermal power plant equipment w , the adjusted TF-IDF calculation formula is:

[0093] TF-IDF w,d =α w ×TF w,d ×IDF w

[0094] in For keyword w k The inverse document frequency measures the importance of a word in a document collection. σ is the total number of documents in the text collection, that is, the number of documents involved in the calculation. is the number of documents originally containing the keyword wk, is the number of documents containing the keyword wk after correction; is the number of documents without keyword wu, No document frequency correction is done for non-keywords, i.e. equal

[0095] The time domain analysis method in step S3 is as follows: the electrical signal data extracted from the data set is a discrete time series x[n'], n'=0,1,2,…,N'-1, where N' is the number of signal sampling points and the sampling interval is T s ;

[0096] The signal periodicity is analyzed by the autocorrelation function, and the calculation expression is:

[0097]

[0098] where R xx [m] is the autocorrelation function of the signal's temporal variation, where m is the number of delay points and the value range of m is -(N'-1)≤m≤(N'-1);

[0099] Signal amplitude calculation: The signal amplitude is determined by calculating the peak value of the signal. In a discrete signal, the maximum value x of the signal is found by traversing all sampling points. max and the minimum value x min , the signal amplitude calculation formula is:

[0100] A=max(|x max |,|x min |)

[0101] Where A is the calculated signal amplitude, which reflects the fluctuation range of the signal in the time domain and is used to determine the operating status of the device.

[0102] This application uses the time domain analysis method to process the electrical signal data of the circulating water system of a thermal power plant. This electrical signal is presented in a discrete time series. When analyzing the law of signal change over time, the autocorrelation function is used for calculation. The autocorrelation function under different delay points can be used to gain insight into the periodicity and stationary characteristics of the signal. Potential periodic faults such as unbalanced vibration of rotating parts can be discovered, providing clues for fault diagnosis. The signal amplitude is calculated by traversing the sampling points to find the maximum value, and the amplitude is calculated using a formula. The amplitude intuitively reflects the signal fluctuation range. When the equipment is operating normally and abnormally, the amplitude changes significantly, which can help staff quickly judge the operating status. The time domain analysis method is simple and efficient. Based on the characteristics of discrete electrical signals, it does not require complex transformations and can quickly process data, providing a basis for fault diagnosis and prediction. Combined with other analysis methods, it can clarify the fault details, help formulate maintenance strategies, and ensure stable system operation.

[0103] Calculation of signal duration in step S3: Assume that the signal starts at the starting time From the beginning to the end End, the signal amplitude is greater than the threshold x during this time period th , by traversing the signal sequence, find The first point and the last point Then the signal duration B is:

[0104]

[0105] Where B is the calculated signal duration, which reflects the duration of the equipment abnormal state and is used to assess the severity of the fault. Based on the time domain analysis method, a comprehensive analysis of the electrical signal data is performed to obtain the signal change pattern over time, amplitude and duration, and key electrical signal characteristics.

[0106] When performing time domain analysis on the electrical signal data of the circulating water system of a thermal power plant, the present application calculates the signal duration by traversing the signal sequence, finding the starting and ending points where the amplitude is greater than a threshold, and calculating the duration using a formula. This indicator reflects the duration of the abnormal state of the equipment and is used to assess the severity of the fault. Combined with the previous analysis of the signal change pattern over time and amplitude calculation, the signal duration allows the time domain analysis method to analyze the electrical signal more comprehensively, presenting the equipment operating status from multiple dimensions and helping to discover complex anomalies. This is not only conducive to accurate fault diagnosis, but also assists in formulating reasonable maintenance strategies, determining the timing of equipment repairs, improving maintenance efficiency, and reducing costs. At the same time, this calculation method based on discrete signal sequences adapts to real-time monitoring needs, can quickly respond to equipment anomalies, and ensure stable system operation.

[0107] In step S4, the dimension of the feature vector extracted from the maintenance report text and the feature vector extracted from the electrical signal are checked, and the vector dimensions of the two groups are adjusted based on the zero-padding method to keep the number consistent; the electrical signal features and the text feature vectors are concat- ed on the feature dimension; and through splicing, a comprehensive feature vector of multimodal information is obtained.

[0108] This application will splice the feature vectors extracted from the maintenance report text and the electrical signal, which can organically combine the information of different modes, so that the comprehensive feature vector contains both the real-time operating status information of the equipment reflected by the electrical signal and the historical maintenance-related knowledge in the maintenance report text, realizing the complementarity of multi-source information. The feature vectors of different modes describe the equipment status from different angles. After splicing, they can present the overall situation of the equipment more comprehensively, avoid the one-sidedness of single modality information, and help to more accurately analyze possible problems and failure trends of the equipment.

[0109] The cross-validation process in step S5 includes:

[0110] (a) Divide the sample set into a training set and a validation set using the k-fold cross-validation method;

[0111] (b) Construct a prediction model based on a deep neural network. The network structure includes an input layer, three fully connected hidden layers, and an output layer, and the activation function is ReLU;

[0112] (c) Using the cross entropy loss function as the optimization objective, combined with the Adam optimization algorithm to dynamically adjust the learning rate, and update the network weight parameters through back propagation;

[0113] (d) Repeat k rounds of cross-validation, with k set to 5 and 10, and perform k rounds of cycles. In each round, 1 fold is selected as the validation set, and the remaining k-1 folds are combined into the training set.

[0114] (e) Calculate the average F1-score of each round and select the model parameters with the highest average F1-score as the final model;

[0115] (f) Use the test set to evaluate the model and calculate the evaluation indicators on the test set. If there are differences, the model is readjusted. If there are no differences, the trained prediction model is obtained.

[0116] During the implementation of this application, the data is pre-divided into multiple parts. Each time a cross-validation operation is performed, one of the data is randomly selected as the validation set. This validation set is responsible for testing the performance of the model, while the remaining data are combined and integrated to form the training set. The main task of the training set is to provide rich learning materials for the model and help the model build a mapping relationship between features and faults.

[0117] During the model training phase, the cross-entropy loss function is used as the key optimization indicator. The cross-entropy loss function can effectively measure the degree of difference between the model's prediction results and the true labels. During the training process, by continuously adjusting various parameters within the model, the cross-entropy loss function will adjust the model parameters based on the difference between the prediction results and the true labels through the back-propagation algorithm, so that the model will be closer to the actual situation in subsequent predictions.

[0118] After the model has undergone multiple rounds of parameter adjustment on the training set, so that the cross-entropy loss function value reaches a certain degree of optimization, it will be predicted on the reserved validation set. After the prediction is completed, the accuracy of the model prediction needs to be quantitatively evaluated. The accuracy indicator used here is the calculation method of the accuracy based on the consistency between the model prediction results and the actual fault labels. Specifically, the number of consistent model prediction results and actual fault labels is counted, and then this number is divided by the total number of validation set samples. The resulting ratio is the accuracy rate. The validation set has a total of 100 samples, of which 80 samples accurately predict the fault status, so the accuracy rate is 80%. In this way, we can intuitively understand the prediction performance of the model on the validation set;

[0119] During multiple rounds of cross-validation, each verification will obtain an accuracy evaluation index. Finally, the system will comprehensively consider the evaluation indicators of all cross-validation rounds and select the model parameters with the best average evaluation index in the cross-validation as the final model parameters. The model determined by these optimal parameters performs best in terms of the correlation between learning features and faults, and can reliably predict faults for new samples. When new thermal power plant circulating water system operation data is input, the model can accurately determine whether there is a fault in the equipment and the possible type of fault based on the learned features and fault association patterns, providing strong support for staff to take maintenance measures in advance and ensure stable system operation.

[0120] In step S5, after determining the existence of a fault risk, the model gives a predicted fault type to ensure the stable operation of the thermal power plant's circulating water system; the predicted fault types include pipeline failure, pump wear, pump cavitation, and fan failure.

[0121] The thermal power plant equipment failure prediction system based on multimodal data includes:

[0122] an electrical signal collection module configured to collect sound waves in the circulating water system and convert them into electrical signals;

[0123] A historical maintenance information collection module configured to collect and process historical maintenance reports in the thermal power plant;

[0124] an electrical signal characteristic analysis module, configured to perform characteristic analysis on the electrical signal;

[0125] A fusion module is configured to fuse the extracted electrical signal features with the feature vectors extracted from the maintenance report text based on a splicing method to form a comprehensive feature of multimodal information;

[0126] The prediction module is configured to perform pre-training learning based on the correlation between learning features and faults using the cross-validation method, input the comprehensive features of multimodal information into the model for prediction output, and predict faults in the circulating water system of the thermal power plant.

[0127] Regular expression filtering case

[0128] The thermal power plant has the following historical maintenance report text materials:

[0129] On May 10, 2023, circulating water pump A experienced abnormal vibration. After disassembly and inspection, it was found that the impeller was worn and subsequently replaced. On June 15, 2023, cooling tower fan B malfunctioned and the motor could not start. Inspection revealed a short circuit in the motor winding, so the motor was replaced.

[0130] Equipment name matching pattern: Constructed based on common equipment names in circulating water systems, circulating water pump | cooling tower fan. This expression can match circulating water pump A and cooling tower fan B that appear in the text.

[0131] Maintenance operation matching mode: constructed using common maintenance action vocabulary, such as disassembly|inspection|replacement. In the above text, it can match the maintenance operations of disassembly and inspection, impeller replacement, and motor replacement.

[0132] Fault phenomenon matching mode: Built based on the description of circulating water system fault phenomena, abnormal vibration, motor failure to start, motor winding short circuit, abnormal vibration, motor failure to start, motor winding short circuit in the text can all be matched;

[0133] After splitting the text into lines, each line is matched against the constructed regular expression. For example, the first line, "On May 10, 2023, circulating water pump A experienced abnormal vibration. It was disassembled and inspected, and the impeller was found to be worn. The impeller was subsequently replaced." This line matches the equipment name matching pattern, circulating water pump A, maintenance operation matching pattern, disassembly inspection, and impeller replacement, and the fault phenomenon matching pattern. Therefore, this line is filtered out. Similarly, the second line of text will also be filtered out because it matches the corresponding pattern.

[0134] Extract matching lines and context and classify them

[0135] For the filtered rows, the context is extracted around them to form information fragments. Assuming that the matching row is the center, one row is extracted before and after. For the first line of text, the extracted information fragment is "On May 10, 2023, circulating water pump A experienced abnormal vibration. It was disassembled and inspected, and the impeller was found to be worn. The impeller was subsequently replaced. All such information fragments are collected and classified according to equipment type, fault type, etc. to form a maintenance information dataset."

[0136] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. Various changes and improvements are possible without departing from the spirit and scope of the present invention, and such changes and improvements fall within the scope of the invention as claimed.

Claims

1. A method for predicting equipment failure in thermal power plants based on multimodal data, characterized in that: The prediction steps are: S1. An acoustic wave transmitter emits acoustic waves in the circulating water system of a thermal power plant, and a sensor receives the emitted elastic waves, converts them into electrical signals, and stores the electrical signals after preprocessing. S2. Based on historical maintenance reports from the thermal power plant, obtain text information, filter out maintenance information related to the circulating water system from the text information using regular expressions, and obtain text feature vectors from the maintenance information using the TF-IDF algorithm; S3. Analyze the electrical signal based on the time domain analysis method to obtain the law of signal change over time, and calculate the signal amplitude and signal duration to obtain the electrical signal characteristics; S4, fusing the electrical signal features with the text feature vector based on the splicing method to form a multimodal comprehensive feature; S5. Build a prediction model and pre-train the correlation between learning features and faults based on the cross-validation method; input multimodal comprehensive features into the prediction model, predict faults in the circulating water system of the thermal power plant, and output the prediction results; The time domain analysis method in step S3 converts the electrical signal into a discrete time series and calculates and analyzes it through the autocorrelation function to obtain the law of signal change over time; the signal amplitude is obtained by calculating the peak value of the signal, and the signal duration calculation reflects the duration of the abnormal state of the equipment.

2. The method for predicting equipment failure in a thermal power plant based on multimodal data according to claim 1, characterized in that: In step S2, text information is obtained based on the equipment operation records and maintenance report text data accumulated by the thermal power plant, and regular expressions are used to filter the text information. The filtering steps are: matching by scanning the text information line by line, extracting the context of the matching lines, and obtaining maintenance information related to the circulating water system.

3. The method for predicting equipment failure in a thermal power plant based on multimodal data according to claim 2, characterized in that: Regular expression construction includes: equipment name matching mode, maintenance operation matching mode and fault phenomenon matching mode. The equipment name matching mode uses the equipment name of the circulating water system as the keyword to construct the expression; the maintenance operation matching mode uses the maintenance action as the keyword to construct the expression; the fault phenomenon matching mode uses the fault phenomenon description as the keyword to construct the expression; Line-by-line scanning matches each piece of text information obtained, splits the text information into lines, matches it line by line with the constructed regular expression, and uses the successfully matched text information as maintenance information related to the circulating water system.

4. The method for predicting equipment failure in a thermal power plant based on multimodal data according to claim 1, characterized in that: The steps based on the TF-IDF algorithm in step S2 are: TF-IDF calculation: Use natural language processing tools to segment the selected maintenance information, split the text into individual words, count the number of times each word appears in the text, obtain the original word frequency, and normalize the word frequency using the following formula: Among them TF ij Indicates the frequency of the jth word in the i-th text, n ij is the number of occurrences of the jth word in the i-th text, is the sum of the number of times all words appear in the i-th text; Determine the inverse document frequency: traverse the entire maintenance information text collection, count the number of documents containing each word, and calculate the inverse document frequency; The collection of words related to thermal power plant equipment is V e , for each vocabulary v∈V in the set e , assign weighting factor α v , where α v ≥1, for non-thermal power plant equipment related words, the weighting factor α v =1; Introducing the thermal power plant equipment dictionary D for the keyword w k , w k ∈D, adjust the document frequency statistics and introduce a correction factor β, then the corrected document frequency is For non-keyword w u ,Right now The document frequency remains constant, i.e. For the keyword w k The inverse document frequency calculation formula is: For non-keyword w n The inverse document frequency calculation formula is: Combined with the weighting factor α of the vocabulary related to thermal power plant equipment w , the adjusted TF-IDF calculation formula is: TF-IDF w,d =α w ×TF w,d ×IDF w in For keyword w k The inverse document frequency measures the importance of a word in a document collection. σ is the total number of documents in the text collection, that is, the number of documents involved in the calculation. is the number of documents originally containing the keyword wk, is the number of documents containing the keyword wk after correction; is the number of documents without keyword wu, No document frequency correction is done for non-keywords, i.e. equal 5. The method for predicting equipment failure in a thermal power plant based on multimodal data according to claim 1, characterized in that: The steps of the time domain analysis method in step S3 are: The electrical signal is recorded as a discrete time series x[n'], n'=0,1,2,…,N'-1, where N' is the number of signal sampling points and the sampling interval is T s ; The signal periodicity is analyzed by the autocorrelation function, and the calculation expression is: where R xx [m] is the autocorrelation function of the signal's temporal variation, where m is the number of delay points and the value range of m is -(N'-1)≤m≤(N'-1); Signal amplitude calculation: The signal amplitude is determined by calculating the peak value of the signal. In a discrete signal, the maximum value x of the signal is determined by traversing all sampling points. max and the minimum value x min , the signal amplitude calculation formula is: A=max(|x max |,|x min |) Where A is the calculated signal amplitude.

6. The method for predicting equipment failure in a thermal power plant based on multimodal data according to claim 1, characterized in that: Calculation of signal duration in step S3: Assume that the signal starts at the starting time From the beginning to the end End, the signal amplitude is greater than the threshold x during this time period th , by traversing the signal sequence, find The first point and the last point The calculation formula based on the duration B of continuous signal interruption is: Where B is the calculated signal duration.

7. The method for predicting equipment failure in a thermal power plant based on multimodal data according to claim 1, characterized in that: The concatenation method in step S4 is as follows: perform dimensionality check on the feature vectors extracted from the maintenance report text and the feature vectors extracted from the electrical signal, adjust the vector dimensions of the two groups based on the zero-padding method to keep the number consistent; perform a concat operation on the electrical signal features and the text feature vectors in the feature dimension to obtain a comprehensive feature vector of multimodal information.

8. The method for predicting equipment failure in a thermal power plant based on multimodal data according to claim 1, characterized in that: The cross-validation process in step S5 includes: (a) Divide the sample set into a training set and a validation set using the k-fold cross-validation method; (b) Construct a prediction model based on a deep neural network. The network structure includes an input layer, three fully connected hidden layers, and an output layer, and the activation function is ReLU; (c) Using the cross entropy loss function as the optimization objective, combined with the Adam optimization algorithm to dynamically adjust the learning rate, and update the network weight parameters through back propagation; (d) Repeat k rounds of cross-validation, with k set to 5 and 10, and perform k rounds of cycles. In each round, 1 fold is selected as the validation set, and the remaining k-1 folds are combined into the training set. (e) Calculate the average F1-score of each round and select the model parameters with the highest average F1-score as the final model; (f) Use the test set to evaluate the model and calculate the evaluation indicators on the test set. If there are differences, the model is readjusted. If there are no differences, the trained prediction model is obtained.

9. The method for predicting equipment failure in a thermal power plant based on multimodal data according to claim 1, characterized in that: In step S5, after determining the existence of a fault risk, the model gives a predicted fault type to ensure the stable operation of the thermal power plant's circulating water system; the predicted fault types include pipeline failure, pump wear, pump cavitation, and fan failure.

10. A thermal power plant equipment failure prediction system based on multimodal data, characterized in that: include: an electrical signal collection module configured to collect sound waves in the circulating water system and convert them into electrical signals; A historical maintenance information collection module configured to collect and process historical maintenance reports in the thermal power plant; an electrical signal characteristic analysis module, configured to perform characteristic analysis on the electrical signal; A fusion module is configured to fuse the extracted electrical signal features with the feature vectors extracted from the maintenance report text based on a splicing method to form a comprehensive feature of multimodal information; The prediction module is configured to perform pre-training learning based on the correlation between learning features and faults using the cross-validation method, input the comprehensive features of multimodal information into the model for prediction output, and predict faults in the circulating water system of the thermal power plant.

Images