Heat energy storage system pipeline vibration abnormity identification method

Abstract

The invention provides a thermal energy storage system pipeline vibration anomaly identification method, and belongs to the technical field of thermal energy storage systems. Vibration sensors are installed at key nodes of a thermal energy storage system pipe network, a topology model based on a graph theory is constructed, and wavelet packet decomposition and empirical mode decomposition technologies are adopted to extract multi-dimensional vibration feature vectors; a working condition identification code system is established, a longest common subsequence algorithm is utilized to match a historical working condition reference, an upper-layer game model with detection precision maximization as a target and a lower-layer game model with a false alarm rate minimization as a target are constructed to realize parameter collaborative optimization, and accurate positioning of an abnormal source is realized by combining finite element transfer function calculation. And according to the deviation measurement value, grading anomaly judgment is performed, and a self-adaptive monitoring frequency adjustment and sliding window database updating mechanism is adopted, so that the technical problems of insufficient pipeline vibration anomaly detection precision and relatively high false alarm rate of the thermal energy storage system are solved.

Description

Technical Field

[0001] This invention belongs to the technical field of thermal energy storage systems, and more specifically, relates to a method for identifying abnormal vibrations in pipelines of thermal energy storage systems. Background Technology

[0002] Vibration monitoring of pipelines in thermal energy storage systems is a crucial technical aspect for ensuring the safe and stable operation of the system. Traditional vibration anomaly detection methods mainly employ single-dimensional signal processing techniques such as fixed threshold comparison and spectrum analysis. These methods involve deploying vibration sensors in key pipe sections to collect vibration signals in real time and comparing them with preset thresholds; when the vibration amplitude exceeds the set threshold, an abnormal state is identified. These methods are widely used in industrial pipeline monitoring and vibration diagnosis of petrochemical plants, offering advantages such as simplicity and low cost. However, traditional vibration detection technologies have significant drawbacks. Due to the complex and variable operating conditions of thermal energy storage systems, the normality of the same vibration signal varies significantly under different conditions. Fixed threshold methods struggle to adapt to these changes and are prone to false alarms. Furthermore, traditional methods often rely on single characteristic parameters for anomaly judgment, lacking comprehensive analysis of the multi-dimensional characteristics of vibration signals, resulting in limited detection accuracy. Additionally, existing technologies lack effective anomaly source localization mechanisms, making it difficult to quickly and accurately determine the location of anomalies. Therefore, current technologies suffer from insufficient accuracy and a high false alarm rate in thermal energy storage system pipeline vibration anomaly detection due to the lack of adaptive anomaly detection mechanisms for multi-condition changes and comprehensive analysis methods for multi-dimensional vibration characteristics. Summary of the Invention

[0003] In view of this, the present invention provides a method for identifying abnormal vibration in pipelines of thermal energy storage systems, which can solve the technical problems of insufficient detection accuracy and high false alarm rate in existing technologies for abnormal vibration in pipelines of thermal energy storage systems.

[0004] This invention is implemented as follows: A method for identifying abnormal vibrations in pipelines of a thermal energy storage system is provided. Vibration sensors are installed at key nodes of each pipe section in the thermal energy storage system, and a pipeline topology model based on graph theory is established. The original vibration signals from each sensor are collected and denoised using wavelet packet decomposition and adaptive spectral kurtosis selection. Empirical mode decomposition is used to extract the intrinsic mode functions of the denoised vibration signals, forming a multidimensional vibration feature vector. An operating condition identification code is established based on the current system operating parameters, and the longest common subsequence algorithm is used to find historical operating condition benchmarks in a multi-operating condition vibration fingerprint database. An upper-level game model and a lower-level game model are established, and the optimal parameters of the vibration anomaly identification model are determined through two-level game optimization. The current multidimensional vibration feature vector is input into the vibration anomaly identification model to calculate the anomaly probability distribution and anomaly type prediction results for each node in the pipeline network. The anomaly source location is located based on the finite element transfer function and the shortest path algorithm. The deviation metric between the current anomaly probability distribution and the historical operating condition benchmark is calculated for anomaly judgment. The vibration monitoring frequency is adjusted according to the anomaly type and severity. The multi-operating condition vibration fingerprint database is updated.

[0005] Specifically, the graph-based pipeline topology model abstracts the thermal energy storage system pipeline network into a graph structure, with pipe segments as edges and connection points as nodes, recording the connection relationships between pipe segments and the pipeline geometric parameters.

[0006] The upper-level game model aims to maximize the vibration anomaly detection accuracy, while the lower-level game model aims to minimize the false alarm rate. The objective functions are the detection accuracy function and the false alarm rate function, and the two objective functions are coupled through an anomaly discrimination threshold.

[0007] Specifically, the adaptive spectral kurtosis selection is a method that automatically determines the optimal filtering frequency band based on the statistical characteristics of the signal. It selects the frequency range containing the most fault information by calculating the spectral kurtosis value of each frequency band.

[0008] Specifically, the vibration anomaly identification model is an anomaly detection model built based on deep learning technology. It takes a multi-dimensional vibration feature vector as input and outputs anomaly probability distribution and anomaly type prediction results.

[0009] Specifically, the multidimensional vibration feature vector is a multidimensional array containing frequency domain, time domain, and energy characteristics of the vibration signal, used to describe the comprehensive characteristics of the vibration signal.

[0010] The operating condition identification code is specifically a character sequence generated according to a fixed coding rule based on the current load level, flow rate, and temperature operating parameters of the thermal energy storage system, and is used to identify different operating conditions.

[0011] Specifically, the multi-condition vibration fingerprint database is a knowledge base that stores the normal vibration characteristics under different conditions, including the vibration spectrum and amplitude distribution standard patterns of each condition.

[0012] Specifically, the finite element transfer function describes the propagation characteristics of vibration signals in the pipeline structure and is used to calculate the propagation path and attenuation coefficient of vibration in the pipeline network.

[0013] The deviation metric is specifically a quantitative indicator of the difference between the current abnormal probability distribution and the historical operating condition benchmark. It is calculated using Euclidean distance. When the deviation metric is ∈ [0, 0.25), it is judged to be in a normal state. When the deviation metric is ∈ [0.25, 0.7), a secondary detection process is triggered. When the deviation metric is ∈ [0.7, 1], the existence of vibration abnormality is confirmed.

[0014] Specifically, the wavelet packet decomposition technique decomposes the vibration signal simultaneously in the time and frequency domains to obtain signal components in multiple frequency sub-bands, which are used to extract multi-resolution vibration features. The empirical mode decomposition is an adaptive signal processing method that decomposes a nonlinear, non-stationary signal into several eigenmode functions, calculating the energy distribution and phase characteristics of each frequency component. The longest common subsequence algorithm is a dynamic programming method used to find the historical operating condition identification code most similar to the current operating condition identification code, achieving similarity matching by constructing a two-dimensional dynamic programming table. The shortest path algorithm is a graph theory algorithm used to find the shortest distance between two points, combined with anomaly probability distribution to determine the shortest path for vibration propagation and locate the anomaly source.

[0015] Specifically, the abnormal probability distribution refers to the probability value of each pipeline node experiencing an anomaly, output by the vibration anomaly identification model. The value ranges from 0 to 1, with a higher value indicating a higher probability of an anomaly.

[0016] Specifically, the historical operating condition benchmark is a reference standard for historical vibration characteristics that is most similar to the current operating condition, obtained by matching from a multi-condition vibration fingerprint database. The system operating condition parameters include the load level, flow rate, temperature, pressure, heat charging and discharging mode, state of the heat storage medium, pipeline flow velocity, and system start-up and shutdown status parameters of the thermal energy storage system.

[0017] This invention effectively solves the technical problems of insufficient detection accuracy and high false alarm rate of traditional methods by constructing a pipeline topology model based on graph theory and a multi-condition vibration fingerprint database, combined with a vibration anomaly identification model optimized by a two-level game theory approach. By employing a multi-dimensional feature extraction technique combining wavelet packet decomposition and empirical mode decomposition, it can fully mine the feature information of vibration signals in multiple dimensions such as time domain, frequency domain, and energy distribution, significantly improving the accuracy of anomaly detection compared to single feature parameters. By establishing a matching mechanism of condition identification code and longest common subsequence algorithm, it achieves adaptive selection of vibration benchmarks under different conditions, effectively avoiding false alarms caused by changes in conditions. By constructing an upper-level game theory model aimed at maximizing detection accuracy and a lower-level game theory model aimed at minimizing the false alarm rate, it achieves synergistic optimization of detection accuracy and false alarm rate, significantly reducing the false alarm rate while improving detection accuracy. In summary, this invention solves the technical problems of insufficient detection accuracy and high false alarm rate in pipeline vibration anomaly detection in thermal energy storage systems mentioned in the background art. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention.

[0019] Figure 2 This is a diagram of the pipeline topology of the thermal energy storage system in the embodiment.

[0020] Figure 3 This is a probability distribution diagram of abnormal pipeline nodes in the embodiment. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0022] like Figure 1 The diagram shows a flowchart of a method for identifying abnormal vibrations in pipelines of a thermal energy storage system provided by the present invention. This method includes the following steps: S01. Install vibration sensors at key nodes of each pipe section in the thermal energy storage system, establish a pipe network topology model based on graph theory, and record the connection relationships and pipe geometric parameters between each pipe section. S02. Collect the original vibration signals from each vibration sensor, decompose the vibration signals into multiple frequency sub-bands using wavelet packet decomposition technology, and use spectral kurtosis adaptive selection of the optimal filtering frequency band for noise reduction processing. S03. The intrinsic mode functions of the noise-reduced vibration signal are extracted by empirical mode decomposition, and the energy distribution and phase characteristics of each frequency component are calculated to form a multidimensional vibration feature vector. S04. Establish a working condition identification code based on the current system working condition parameters, and use the longest common subsequence algorithm to find the historical working condition benchmark that is most similar to the current working condition identification code in the multi-working condition vibration fingerprint database. S05. Establish an upper-level game model aimed at maximizing the accuracy of vibration anomaly detection and a lower-level game model aimed at minimizing the false alarm rate. Optimal parameters of the vibration anomaly identification model are determined through two-level game optimization. S06. Input the current multidimensional vibration feature vector into the vibration anomaly identification model to calculate the anomaly probability distribution and anomaly type prediction results of each node in the pipeline network. S07. Calculate the propagation path and attenuation coefficient of vibration in the pipeline based on the finite element transfer function, and use the shortest path algorithm to locate the source of the anomaly by combining the anomaly probability distribution. S08. Calculate the deviation metric value between the current abnormal probability distribution and the historical working condition benchmark. When the deviation metric value ∈ [0, 0.25), it is judged as a normal state. When the deviation metric value ∈ [0.25, 0.7), the secondary detection process is triggered. When the deviation metric value ∈ [0.7, 1], the vibration abnormality is confirmed. S09. Adjust the vibration monitoring frequency according to the type and severity of the abnormality. Increase the monitoring frequency by 30% for minor abnormalities, increase the monitoring frequency by 60% for moderate abnormalities, and start the continuous monitoring mode for severe abnormalities. S10. Optional, update the multi-condition vibration fingerprint database by adding the multi-dimensional vibration feature vectors under the current condition to the corresponding condition category and using a sliding window mechanism to keep the database size stable.

[0023] Among them, the graph-based pipeline topology model abstracts the thermal energy storage system pipeline network into a graph structure, with pipe segments as edges and connection points as nodes, recording the geometric parameters such as the length, diameter, and material of each pipe segment, as well as the connection relationships between nodes. Wavelet packet decomposition technology decomposes the vibration signal simultaneously in the time and frequency domains to obtain signal components at different frequencies and time windows, which are used to extract multi-resolution vibration features.

[0024] Spectral kurtosis adaptive selection is a method that automatically determines the optimal filtering frequency band based on the statistical characteristics of the signal. It selects the frequency range containing the most fault information by calculating the spectral kurtosis values ​​of each band. Empirical mode decomposition (EMD) is an adaptive signal processing method that decomposes a nonlinear, non-stationary signal into several eigenmode functions, each representing an inherent vibration mode of the signal. Multidimensional vibration feature vectors are multidimensional arrays containing frequency domain, time domain, and energy characteristics of the vibration signal, used to describe the comprehensive characteristics of the vibration signal.

[0025] Operating condition identification codes are character sequences generated according to fixed coding rules based on the current load level, flow rate, temperature, and other operating parameters of the thermal energy storage system, used to identify different operating conditions. The longest common subsequence algorithm, a dynamic programming method, is used to find the longest common subsequence between two character sequences. A two-dimensional dynamic programming table is constructed to achieve similarity matching between the current operating condition identification code and historical operating condition identification codes. The multi-operating condition vibration fingerprint database is a knowledge base storing normal vibration characteristics under different operating conditions, including standard patterns such as vibration spectrum and amplitude distribution for each operating condition. The historical operating condition benchmark is a reference standard of historical vibration characteristics most similar to the current operating condition, obtained by matching from the multi-operating condition vibration fingerprint database.

[0026] The upper-level game theory model aims to maximize the accuracy of vibration anomaly detection, with the objective function being the detection accuracy function. The lower-level game theory model aims to minimize the false alarm rate, with the objective function being the false alarm rate function. The two objective functions are coupled through an anomaly discrimination threshold. The vibration anomaly identification model is an anomaly detection model built based on deep learning technology. It takes a multi-dimensional vibration feature vector as input and outputs anomaly probability distribution and anomaly type prediction results. The anomaly probability distribution is the probability value of anomalies occurring at each pipeline node output by the vibration anomaly identification model, with values ​​ranging from 0 to 1. A higher value indicates a higher probability of an anomaly. The anomaly type prediction result is the anomaly category identifier output by the vibration anomaly identification model, used to distinguish different types of vibration anomalies.

[0027] The finite element transfer function describes the propagation characteristics of vibration signals in a pipeline structure, including propagation velocity, attenuation law, and frequency response characteristics. The shortest path algorithm, a graph theory algorithm used to find the shortest distance between two points, is used in anomaly source localization to determine the shortest path for vibration propagation. The deviation metric is a quantitative indicator of the difference between the current anomaly probability distribution and historical operating condition benchmarks, calculated using Euclidean distance. The sliding window mechanism is a database maintenance strategy that deletes the oldest data when new data is added, keeping the database capacity within a preset range.

[0028] The detection accuracy function quantifies the accuracy of vibration anomaly detection. Its inputs include the number of true anomalies detected, the total number of anomalies, and a detection threshold parameter; the output is the detection accuracy value. The false alarm rate function quantifies the probability that a normal state is mistakenly identified as an anomaly. Its inputs include the number of false alarms, the total number of normal samples, and a detection threshold parameter; the output is the false alarm rate value. The anomaly discrimination threshold is a coupling parameter connecting the upper-level game model and the lower-level game model, affecting both the detection accuracy and the false alarm rate calculation results.

[0029] The objective function of the upper-level game model is expressed as follows: ,in This represents the actual number of anomalies detected. This represents the total number of anomalies. To detect threshold parameters, This serves as a standard threshold reference value. The objective function of the lower-level game model is expressed as follows: ,in The number of false alarms. This represents the total number of normal samples.

[0030] The specific structure of the vibration anomaly recognition model is a Transformer encoder based on a multi-head attention mechanism, including an input embedding layer, multiple encoder blocks, and a classification output layer. The number of heads in the multi-head attention mechanism is determined based on the dimension of the multi-dimensional vibration feature vector and the number of pipeline nodes. The training dataset establishment steps of the vibration anomaly recognition model specifically include collecting normal and abnormal vibration data under different working conditions, preprocessing the vibration signals to extract multi-dimensional vibration feature vectors, labeling the anomaly type and severity, and dividing the training set, validation set, and test set in an 8:1:1 ratio. The training steps of the vibration anomaly recognition model specifically include using the Adam optimizer for gradient descent, setting the learning rate to 0.001, the batch size to 64, the number of training epochs to 200, and using the cross-entropy loss function and early stopping mechanism to prevent overfitting.

[0031] Furthermore, the present invention is implemented by a computer to form a thermal energy storage system pipeline vibration anomaly identification system. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they execute the above-mentioned method.

[0032] The specific implementation methods of the above steps are described in detail below.

[0033] The specific implementation of step S01 involves mathematically abstracting and modeling the thermal energy storage system pipeline network using graph theory. First, a field survey of the entire thermal energy storage system is conducted to determine the spatial layout and connection relationships of all pipe segments. Then, each pipe segment is abstracted as an edge in the graph, and the connection points of the pipe segments are abstracted as nodes, forming an undirected connected graph structure. During the modeling process, the geometric parameters of each pipe segment, including length, inner diameter, outer diameter, wall thickness, material density, elastic modulus, and other physical property parameters, need to be measured and recorded. Simultaneously, the coordinate position of each node and the number of connected pipe segments are recorded. An adjacency matrix is ​​established to represent the connection relationships between nodes; an element value of 1 in the matrix indicates a direct connection between corresponding nodes, while a value of 0 indicates no direct connection. Based on the actual scale of the pipeline network, an appropriate number of sensors and their installation locations are selected. Typically, one vibration sensor is installed every 50 meters of pipe segment. Sensors must be installed at critical nodes such as pipeline branches, elbows, and tees. During sensor installation, it is necessary to ensure close contact with the pipe surface, and the installation torque should be controlled within the range of 5–10 Nm.

[0034] The specific implementation of step S02 involves synchronously acquiring the raw vibration signals from each vibration sensor using a multi-channel data acquisition system. The sampling frequency is set to 2048 Hz, the data bit depth is 16 bits, and the continuous acquisition time is 60 seconds. The acquired raw signals are first analyzed using wavelet packet decomposition technology for multi-resolution analysis. A biorthogonal wavelet is selected as the mother wavelet function, and the decomposition level is set to 6 levels, decomposing the vibration signal into 64 frequency sub-bands. For each frequency sub-band, a spectral kurtosis index is calculated. Spectral kurtosis is defined as the ratio of the fourth-order cumulant to the square of the second-order cumulant. By comparing the spectral kurtosis values ​​of each sub-band, the top 10 frequency sub-bands with the highest spectral kurtosis values ​​are selected as the optimal filtering frequency bands. A Butterworth bandpass filter is used to filter the selected frequency sub-bands. The filter order is set to 8, the passband ripple is less than 0.1 dB, and the stopband attenuation is greater than 60 dB. The filtered signal is further processed by an adaptive threshold noise reduction algorithm to remove noise interference. The threshold parameter is dynamically adjusted according to the signal's standard deviation, typically set to 1.5 times the signal's standard deviation.

[0035] The specific implementation of step S03 involves using an empirical mode decomposition (EMD) algorithm to adaptively decompose the denoised vibration signal. First, the signal is sieved to identify all local maxima and minima. Then, cubic spline interpolation is used to fit the upper and lower envelopes, and the mean of the upper and lower envelopes is calculated as the sieved signal. The original signal is subtracted from the sieved signal to obtain candidate intrinsic mode functions (EMFs). This sieved process is repeated until the criteria for EMFs are met. These criteria include that the number of extrema equals or differs from the number of zero-crossings by no more than one, and that the mean of the upper and lower envelopes is close to zero across the entire data range. EMFs and a residual component can be obtained through EMFs, typically resulting in 8–12 EMFs. For each EMF, its frequency domain characteristics, including center frequency, bandwidth, and energy contribution ratio, are calculated. Simultaneously, its time domain characteristics, including root mean square (RMS), kurtosis, and skewness, are also calculated. All feature parameters are combined to form a multidimensional vibration feature vector, which typically has a dimension of 50 to 80. Each element in the vector represents a feature attribute of the vibration signal.

[0036] The specific implementation of step S04 involves constructing an operating condition identification code based on the real-time operating status of the thermal energy storage system. First, key operating parameters of the system are collected, including inlet water temperature, outlet water temperature, circulation flow rate, system pressure, and load power. These parameters are then quantified and encoded. Temperature parameters are classified in 10°C intervals, flow rate parameters in 10% intervals of rated flow rate, and pressure parameters in 5% intervals of rated pressure. Each parameter level is encoded as a single character, and these characters are combined to form the operating condition identification code string. A multi-operating condition vibration fingerprint database stores identification codes and corresponding vibration characteristic reference standards for various historical operating conditions. The database typically contains vibration fingerprint data for 500 to 1000 different operating conditions. The longest common subsequence algorithm is used to calculate the similarity between the current operating condition identification code and the historical operating condition identification codes in the database. The algorithm solves this step-by-step by constructing a two-dimensional dynamic programming table, where each element represents the length of the longest common subsequence of the first few characters of two strings. The similarity calculation formula is the ratio of the longest common subsequence length to the length of the longer string. The historical operating condition with the highest similarity (greater than 0.8) is selected as the benchmark reference for the current operating condition.

[0037] The specific implementation of step S05 involves establishing a two-layer game optimization model to determine the optimal parameter configuration for vibration anomaly identification. The upper-layer game model aims to maximize detection accuracy, while the lower-layer game model aims to minimize the false alarm rate. The two models are coupled through anomaly discrimination threshold parameters. The objective function input parameters of the upper-layer game model include the number of true anomalies detected, the total number of anomalies, the detection threshold parameter, and the standard threshold reference value. The detection accuracy value is calculated using logarithmic functions and ratio operations. The objective function input parameters of the lower-layer game model include the number of false alarms, the total number of normal samples, and the detection threshold parameter. The false alarm rate value is calculated using a combination of exponential and logarithmic functions. The two-layer game model is solved using a particle swarm optimization algorithm. The particle swarm size is set to 50 particles, the number of iterations is set to 200, the inertia weight decreases linearly from 0.9 to 0.4, and the acceleration coefficient is set to 2.0. During the optimization process, the upper-level model and the lower-level model alternately update the parameters. After each iteration, the convergence of the objective function of the two layers is checked. When the change of the objective function is less than 0.001 for 10 consecutive iterations, it is judged as convergence.

[0038] The specific implementation of step S06 involves inputting the multi-dimensional vibration feature vector extracted under the current operating condition into a pre-trained vibration anomaly recognition model for anomaly detection. This model employs a converter encoder structure based on a multi-head attention mechanism. The input embedding layer maps the multi-dimensional vibration feature vector to a 512-dimensional hidden space, and then performs feature learning through eight encoder blocks. Each encoder block contains a multi-head self-attention mechanism and a feedforward neural network. The number of heads in the multi-head attention mechanism is set to eight, and the hidden layer dimension of the feedforward network is 2048 dimensions. During model training, the Adam optimizer is used for parameter updates, with a learning rate set to 0.001, a batch size of 64 samples, and a training cycle of 200 rounds. The cross-entropy loss function is used to calculate the difference between the model's prediction results and the true labels. The model output layer produces two types of results: one is the anomaly probability distribution of each pipeline node, with a value ranging from 0 to 1, representing the likelihood of vibration anomalies occurring at each node; the other is the anomaly type prediction result, classifying anomalies into five categories: slight vibration, moderate vibration, severe vibration, resonance anomaly, and impact anomaly.

[0039] The specific implementation of step S07 is based on the finite element analysis method to calculate the propagation characteristics of vibration signals in the pipeline structure, establish a finite element model of pipeline vibration, and discretize the pipeline into several beam elements, each with 6 degrees of freedom, including 3 translational and 3 rotational degrees of freedom. Based on the pipeline's geometric parameters and material properties, the element stiffness matrix and mass matrix are established, and assembled to obtain the overall stiffness matrix and mass matrix of the entire pipeline network. The modal characteristics of the pipeline network, including natural frequencies and mode shapes, are solved, and the transfer function of vibration at different frequencies is calculated. The transfer function describes the amplitude attenuation and phase delay relationship of vibration from the excitation point to the response point. Combining the anomaly probability distribution results, the Dijkstra shortest path algorithm is used to find the shortest path for vibration propagation in the pipeline network topology. The algorithm starts from the node with the highest anomaly probability and targets each sensor node, calculating the shortest propagation distance. Considering the attenuation effect of vibration in the pipeline, the path weight is adjusted according to the amplitude attenuation characteristics of the transfer function, finally determining the most likely location of the anomaly source, with a positioning accuracy typically within 5 meters.

[0040] The specific implementation of step S08 involves calculating the deviation metric between the current anomaly probability distribution and the historical operating condition benchmark to determine the abnormal state of the system. The Euclidean distance calculation method is used, and the deviation metric is obtained by taking the square root of the sum of the squares of the differences between the current anomaly probability values ​​of each node and the corresponding historical benchmark values. Three judgment intervals are set for state classification. When the deviation metric is within the range of 0 to 0.25, the system is considered to be in normal operation, at which point the anomaly probabilities of each node are within the normal fluctuation range, and the system is operating stably. When the deviation metric is within the range of 0.25 to 0.7, a secondary detection process is triggered, increasing the sampling frequency and extending the monitoring time, recalculating the vibration characteristics and anomaly probabilities, and confirming the abnormal state through the consistency of three consecutive detection results. When the deviation metric is within the range of 0.7 to 1.0, a vibration anomaly is directly confirmed, at which point the system's vibration mode differs significantly from the normal operating condition, requiring immediate initiation of the anomaly handling procedure. In practical applications, these thresholds can be fine-tuned according to the specific characteristics of the thermal energy storage system and the operating environment. Typically, the upper limit of the normal range can be adjusted to 0.2 to 0.3, and the lower limit of the abnormal range can be adjusted to 0.6 to 0.8.

[0041] The specific implementation of step S09 involves dynamically adjusting the sampling frequency and monitoring strategy of the vibration monitoring system based on the detected anomaly type and severity, establishing an adaptive monitoring frequency adjustment mechanism. When a minor vibration anomaly is detected, the original monitoring frequency is increased by 30%, for example, from once per hour to once every 46 minutes. After 24 hours of continuous monitoring, if the anomaly disappears, the normal frequency is restored; if the anomaly persists, the monitoring frequency is further increased. When a moderate vibration anomaly is detected, the monitoring frequency is increased by 60%, and a multi-sensor collaborative monitoring mode is activated, increasing the data acquisition frequency of surrounding adjacent sensors and extending the monitoring duration to 72 hours. During this period, the changing trend of the anomaly state is assessed every 6 hours. When a severe vibration anomaly or resonance phenomenon is detected, a continuous monitoring mode is immediately activated, the sampling interval is shortened to 30 seconds, and all relevant sensors are activated for full-network monitoring, continuing monitoring until the anomaly is completely eliminated or corresponding engineering measures are taken. The adjustment of the monitoring frequency needs to consider the limitations of the system's computing resources and storage capacity to avoid overloading the system due to excessive monitoring.

[0042] The specific implementation of step S10 involves using a sliding window mechanism to dynamically update and maintain the multi-condition vibration fingerprint database. When a new combination of conditions is encountered during system operation, the multi-dimensional vibration feature vector and condition identification code under the current condition are added to the corresponding condition category in the database. The database storage capacity is typically set to a maximum of 10,000 records. When the addition of new data exceeds the capacity limit, the oldest data record is deleted using a first-in, first-out (FIFO) principle. During the update process, the newly added vibration feature vectors undergo quality checks, including data integrity verification, outlier detection, and feature vector dimension consistency verification. Only data that passes the quality check is officially added to the database. The database is regularly cleaned and optimized, deleting records with high duplication and merging data from similar conditions to maintain the representativeness and diversity of the database. The database update frequency is set to a batch update every 100 hours of operation, or an update when 50 new condition data entries are accumulated, ensuring that the database reflects the latest changes in the system's operating status.

[0043] It should be noted that the key technical ideas of this invention are mainly reflected in the following aspects. First, the graph theory-based pipeline topology modeling and multi-scale feature fusion technology abstracts the complex thermal energy storage pipeline system into a mathematical graph structure. Combined with multi-scale analysis methods such as wavelet packet decomposition and empirical mode decomposition, it achieves refined feature extraction of pipeline vibration signals. Compared with traditional single-frequency or time-domain analysis methods, this technology can simultaneously capture the feature information of vibration signals at different time and frequency scales, effectively improving the sensitivity and accuracy of anomaly feature identification, especially significantly enhancing the detection capability for early, weak anomaly signals. Second, the innovative application of a two-layer game optimization model. This model achieves intelligent optimization configuration of anomaly detection parameters by constructing an upper-layer game aimed at maximizing detection accuracy and a lower-layer game aimed at minimizing the false alarm rate. Compared with traditional single-objective optimization methods, the two-layer game model can effectively control the false alarm rate while ensuring high detection accuracy, solving the technical problem of balancing detection accuracy and false alarm rate in traditional methods, and significantly improving the practicality and reliability of the anomaly identification system. Thirdly, there is the adaptive benchmark matching and precise anomaly source localization technology. By establishing a benchmark identification code and a multi-benchmark vibration fingerprint database, and combining it with the longest common subsequence algorithm, adaptive benchmark matching under different operating conditions is achieved. Then, the finite element transfer function and the shortest path algorithm are used to achieve precise anomaly source localization. This technology overcomes the poor adaptability problem caused by the fixed detection benchmark under different operating conditions in traditional methods, and at the same time, it elevates anomaly localization from qualitative analysis to quantitative precise localization, providing accurate location information for anomaly handling.

[0044] The synergistic effect of these key technological approaches constructs a complete intelligent vibration anomaly identification system, achieving end-to-end technological integration from signal acquisition, feature extraction, pattern recognition to anomaly localization. Multi-scale feature fusion provides rich and accurate feature information for subsequent anomaly identification, two-layer game optimization ensures optimal performance configuration of the identification algorithm, adaptive matching of operating conditions guarantees the stability of the system under various operating conditions, and precise anomaly localization provides a reliable basis for maintenance decisions. Compared with the current situation where existing technologies mostly adopt single detection methods and lack systematic integration, this invention achieves a comprehensive improvement in detection accuracy, adaptability, and localization accuracy through multi-technology synergy, forming a complete intelligent vibration anomaly identification solution for thermal energy storage systems, and providing a new technical path for health monitoring of complex industrial systems.

[0045] It should be noted that this invention also solves the following technical problem: Existing technologies suffer from inaccurate location of vibration anomaly sources in thermal energy storage system pipelines. While traditional vibration monitoring methods can detect abnormal signals, they lack an effective anomaly source location mechanism. They typically rely on manual experience or simple signal strength comparisons to infer the anomaly location, resulting in limited accuracy and long processing times. This invention constructs a pipeline network topology model based on graph theory, records the geometric parameters and connection relationships of each pipe segment, calculates the vibration propagation path and attenuation coefficient in the pipeline network using the finite element transfer function, determines the optimal path for vibration propagation using the shortest path algorithm, and reversely calculates the anomaly source location based on the anomaly probability distribution of each node. This achieves rapid and accurate anomaly source location, significantly improving fault diagnosis efficiency.

[0046] Furthermore, this invention addresses the technical problem of the lack of adaptability in existing vibration monitoring systems. Traditional monitoring systems typically employ fixed sampling frequencies and data processing modes, failing to dynamically adjust according to system operating status and the severity of anomalies. This leads to resource waste under normal operating conditions and untimely monitoring under abnormal conditions. This invention establishes a tiered anomaly judgment mechanism, automatically adjusting the monitoring strategy based on different ranges of deviation metrics. For minor anomalies, the monitoring frequency is moderately increased; for moderate anomalies, the monitoring density is significantly increased; and for severe anomalies, a continuous monitoring mode is activated. This ensures both the real-time performance and reliability of anomaly detection, while also achieving rational allocation of computing resources, thus improving the overall intelligence and operational efficiency of the monitoring system.

[0047] Specifically, the principle of this invention is as follows: The reason why this invention can solve the technical problem of insufficient accuracy and high false alarm rate in the detection of abnormal vibrations in pipelines of thermal energy storage systems lies in the construction of an intelligent anomaly detection system based on multi-dimensional feature fusion and adaptive operating condition matching. First, the vibration signal is decomposed simultaneously in the time and frequency domains using wavelet packet decomposition technology. Combined with adaptive selection of the optimal filtering frequency band using spectral kurtosis, noise interference can be effectively filtered out while retaining key anomaly feature information. Subsequently, empirical mode decomposition is used to extract the intrinsic mode functions of the signal, calculate the energy distribution and phase characteristics of each frequency component, and form a multi-dimensional feature vector containing rich vibration information, which can more comprehensively describe the vibration state compared with traditional single feature parameters. Second, by establishing an operating condition identification code system and the longest common subsequence algorithm, the accurate matching between the current operating condition and the historical operating condition benchmark is achieved, solving the problem that the traditional fixed threshold method cannot adapt to changes in operating conditions and ensuring the accuracy of the anomaly judgment benchmark. Third, the core of the two-layer game optimization model is to take maximizing detection accuracy and minimizing false alarm rate as mutually coupled optimization objectives. The two objective functions are synergistically optimized through the anomaly discrimination threshold parameter, which mathematically guarantees the optimization of detection performance. Finally, the Transformer encoder based on the multi-head attention mechanism can fully explore the temporal dependencies and spatial correlations in the multi-dimensional vibration feature vectors. Combined with the vibration propagation path information calculated by the finite element transfer function, it can achieve accurate anomaly source localization, thus forming a complete high-precision, low-false-alarm vibration anomaly identification system.

[0048] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0049] In this embodiment, the specific implementation of step S01 is the same as described above, and will not be repeated in detail here.

[0050] The specific implementation of step S02 is to perform multi-resolution analysis of the vibration signal using wavelet packet decomposition technology. The mathematical expression for wavelet packet decomposition is: ; In the formula, For the first Layer The first time window Wavelet packet coefficients of each frequency sub-band; For the first The impulse response coefficients of each filter are calculated using bioorthogonal wavelet functions. The number of decomposition layers, with a value ranging from 1 to 6; Index for time windows; For filter coefficient index; This is the floor function; This is a normalization factor to ensure energy conservation for coefficients of different strata.

[0051] The formula for calculating adaptive spectral kurtosis selection is as follows: ; In the formula, For the first The spectral kurtosis values ​​of each frequency sub-band are dimensionless. For mathematical expectation operators; For the first Wavelet packet coefficients of each frequency sub-band; and These are the fourth and second power moduli of the coefficients, respectively.

[0052] The parameter acquisition method is as follows: The results were obtained by calculating the bioorthogonal wavelet basis functions, using the db4 wavelet with a filter length of 8. The value is determined based on the signal sampling frequency and analysis requirements, and is usually set to 6. and It was obtained through statistical calculations on 60 seconds of acquired signals.

[0053] The specific implementation of step S03 is to extract intrinsic mode functions using the empirical mode decomposition algorithm. The mathematical expression for the screening process is: ; ; In the formula, For the first intrinsic mode function, the first... Secondary screening results; This is the original vibration signal; is the mean function of the upper and lower envelopes; For the first The envelope mean of the second screening; The number of iterations for screening is typically 5 to 15.

[0054] The formula for calculating the energy distribution of intrinsic mode functions is as follows: ; In the formula, For the first The normalized energy contribution ratio of each intrinsic mode function, dimensionless; For the first One intrinsic mode function; The signal length is expressed in units of the number of sampling points. This represents the total number of intrinsic mode functions, typically 8 to 12. This is the index for the intrinsic mode functions.

[0055] The parameter acquisition method is as follows: The local maxima and minima were obtained by fitting the local maxima and minima using cubic spline interpolation. Calculated through a recursive sieving process; the sieving stopping condition is the standard deviation. ,in It is a dimensionless standard deviation index.

[0056] The specific implementation of step S04 is to use the longest common subsequence algorithm to calculate the similarity of working conditions. The dynamic programming recursive formula is expressed as follows: ; The formula for calculating the similarity of operating conditions is as follows: ; In the formula, For the current working condition identification code The character and historical operating condition identification code are before The length of the longest common subsequence of characters; The current working condition identification code number One character; For historical operating condition identification code number One character; The similarity of working conditions is dimensionless and ranges from 0 to 1. and The lengths of the identification codes for the current and historical operating conditions are respectively, typically 8 to 12 characters; and These are the character indices for the current and historical operating condition identification codes, respectively.

[0057] The parameter acquisition method is as follows: and The operating parameters such as temperature, flow rate, and pressure are quantized and coded at intervals of 10℃, 10% of the rated value, and 5% of the rated value. It is obtained by constructing a two-dimensional dynamic programming table and calculating step by step.

[0058] The specific implementation of step S05 is to establish a two-layer game optimization model. The objective function of the upper-layer game model is expressed as follows: ; The objective function of the lower-level game model is expressed as follows: ; The iterative update formula for the game equilibrium solution is expressed as follows: ; In the formula, The actual number of anomalies detected was obtained through verification using a test set. This represents the total number of anomalies. The detection threshold parameter is dimensionless and ranges from 0.1 to 0.9. This is a standard threshold reference value, dimensionless, with an empirical value of 0.5; Number of false alarms; This represents the total number of normal samples; This is the weighting coefficient, dimensionless, with a default value of 0.6; This represents the number of iterations. and All are dimensionless objective functions, and dimensional unification is achieved through a combination of quantity ratios and threshold ratios.

[0059] The parameter acquisition method is as follows: , Obtained through statistics from the confusion matrix; , Statistical analysis of sample labels in the training set; Based on historical experience; The optimal value is determined through cross-validation.

[0060] The specific implementation method of step S06 is the same as described above, and will not be repeated in detail here.

[0061] The specific implementation of step S07 involves calculating the vibration transfer function and locating the anomaly source. The mathematical expression for the transfer function is: ; The shortest path algorithm for anomaly source localization is described as follows: ; In the formula, For frequency The transfer function is dimensionless. and These are the frequency domain representations of the output and input vibration signals, respectively. The imaginary unit; The damping ratio is dimensionless and is experimentally determined; typical values ​​range from 0.02 to 0.05. Angular frequency, unit: ; The natural frequency of the pipeline, in units of ; To reach the node The normalized shortest distance is dimensionless; For nodes arrive Physical distance, in units of ; It is the reciprocal of the magnitude of the transfer function, dimensionless, and used as a path weight correction factor; It is a set of nodes.

[0062] The parameter acquisition method is as follows: Finite element modal analysis was used to calculate the pipeline geometry and boundary conditions. Determined by hammer impact test; Obtained through on-site measurements; and Calculated using Fast Fourier Transform; Normalization is performed by dividing by the maximum distance of the pipeline network to unify the dimensions of each term in the formula.

[0063] The specific implementation of step S08 involves calculating a deviation metric value for anomaly detection. The Euclidean distance calculation formula is as follows: ; In the formula, This is a normalized deviation measure, dimensionless, with a value range of 0 to 1; For the first The current anomaly probability value of each node, dimensionless; For the first The historical baseline probability value of each node is dimensionless. This represents the total number of pipeline nodes. For node indexing.

[0064] The parameter acquisition method is as follows: Obtained through the output of the anomaly detection model; Obtained by matching from a multi-condition vibration fingerprint database; The number of nodes is determined based on the size of the pipeline network, typically ranging from 20 to 100.

[0065] The specific implementation of step S09 is to adjust the monitoring frequency according to the severity of the anomaly. The frequency adjustment formula is expressed as follows: ; In the formula, This is the adjusted monitoring frequency, in units of... ; Basic monitoring frequency, unit: Typically, this is done once per hour; To adjust the coefficient, it is dimensionless, with a value of 0.3 for slight anomalies, 0.6 for moderate anomalies, and continuous monitoring for severe anomalies; The severity level is dimensionless, with values ​​of 1, 2, and 3 corresponding to mild, moderate, and severe, respectively.

[0066] The specific implementation of step S10 is the same as described above, and will not be repeated in detail here.

[0067] The principle of wavelet packet decomposition formula is based on multi-resolution analysis theory. By decomposing the signal simultaneously in the time domain and frequency domain, it achieves refined extraction of different frequency components of the vibration signal.

[0068] ; Compared with the traditional Fourier transform, this formula has better time-frequency localization characteristics, which can effectively identify transient vibration anomaly signals. In particular, it significantly improves the analysis effect of non-stationary vibration signals, providing a richer information basis for subsequent anomaly feature extraction.

[0069] The principle of the spectral kurtosis calculation formula is based on the theory of higher-order statistics. It measures the impulsiveness of a signal by calculating the ratio of the fourth-order cumulant to the second-order cumulant.

[0070] ; This formula can adaptively identify the frequency range containing the most fault information. Compared with the fixed frequency band filtering method, it has stronger adaptability, effectively improves the detection rate of abnormal signals and the accuracy of feature extraction, and reduces the subjectivity of manually setting parameters.

[0071] The principle of the empirical mode decomposition sieving formula is based on the natural separation of the intrinsic vibration modes of the signal. Through a recursive sieving process, the complex nonlinear nonstationary signal is decomposed into several single-component signals.

[0072] ; Compared with traditional frequency domain decomposition methods, this method is completely data-driven, does not require preset basis functions, and can adaptively extract the inherent vibration characteristics of the signal, making it particularly suitable for vibration signal analysis under complex operating conditions of thermal energy storage systems.

[0073] The principle of the energy distribution calculation formula is based on the distribution law of signal energy on different frequency components. Through normalization, the comparability of vibration characteristics under different working conditions is achieved.

[0074] ; This formula provides a quantitative description of the frequency structure of vibration signals. Compared with single amplitude features, it has stronger stability and representativeness, providing a reliable feature basis for anomaly identification.

[0075] The principle of the longest common subsequence algorithm is based on dynamic programming theory. It achieves accurate calculation of the similarity between codes for working condition identification by constructing state transition equations.

[0076] ; Compared with simple string matching methods, this algorithm has stronger fault tolerance, can handle small fluctuations in operating parameters, realizes intelligent matching of vibration benchmarks under different operating conditions, and significantly improves the adaptability of anomaly detection.

[0077] The objective function of the two-layer game model is based on multi-objective optimization theory. By constructing a competitive relationship between maximizing detection accuracy and minimizing false alarm rate, the optimal configuration of anomaly detection parameters is achieved.

[0078] ; ; Compared with single-objective optimization methods, this model can consider multiple aspects of detection performance at the same time, effectively solving the problem of difficulty in balancing accuracy and false alarm rate in traditional methods, and greatly improving the practicality of anomaly recognition systems.

[0079] The principle of the transfer function formula is based on the theory of vibration dynamics, which describes the propagation and attenuation law of vibration signals in pipe structures.

[0080] ; This function takes into account the damping characteristics and frequency response of the pipeline, and has higher physical accuracy than the simple distance attenuation model. It provides a theoretical basis for the accurate location of anomaly sources and significantly improves the location accuracy.

[0081] The principle of the deviation metric calculation formula is based on the multidimensional spatial distance theory, and the normalized Euclidean distance is used to achieve a quantitative assessment of the degree of anomaly.

[0082] ; Compared with the single-point threshold judgment method, this formula has stronger robustness, can comprehensively consider the vibration state changes of the entire pipeline network, effectively reduce the impact of local disturbances on the overall judgment, and improve the reliability of anomaly detection.

[0083] The principle of the monitoring frequency adjustment formula is based on adaptive control theory. By establishing a quantitative relationship between frequency and the severity of anomalies, intelligent optimization of the monitoring strategy is achieved.

[0084] ; Compared with the fixed-frequency monitoring method, this formula has better resource utilization efficiency, can reduce the system's computational load while ensuring monitoring effectiveness, and achieves the optimal balance between monitoring accuracy and system efficiency.

[0085] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2: A certain thermal energy storage system includes a cold salt storage tank, a hot salt storage tank, and a main pipeline system connecting the tanks. The total length of the pipeline network is approximately 850m. The diameter of the main pipeline is 600mm, and the diameters of the branch pipelines are 400mm and 200mm, respectively. The pipeline material is 316L stainless steel, the design temperature is 565℃, and the design pressure is 1.6MPa.

[0086] The technical team first installed high-precision vibration sensors at 32 key nodes of the thermal energy storage system. The sensors are triaxial accelerometers with a measurement range of ±50°. The frequency response range is 0.5–10000Hz, and the sampling frequency is set to 25600Hz. A pipeline network topology model was established based on graph theory, abstracting the pipeline network as a directed graph containing 32 nodes and 47 edges, recording the geometric parameters of each pipe segment, such as length, diameter, wall thickness, and material density. Figure 2 As shown, the established pipeline topology model can clearly reflect the connection relationship between each pipe segment and the vibration propagation path.

[0087] During the vibration signal processing phase, the technical team collected 72 consecutive hours of raw vibration data, totaling 156GB. Wavelet packet decomposition was used to divide the vibration signal into eight frequency sub-bands, ranging from 0.1Hz to 12800Hz. Using a spectral kurtosis adaptive selection algorithm, the optimal filtering band was determined to be 800–3200Hz, which contains over 95% of the effective vibration information. After noise reduction, the signal-to-noise ratio improved from 12.3dB to 28.7dB.

[0088] Empirical Mode Decomposition (EMD) was used to extract the intrinsic mode functions (EMFs) of the denoised vibration signal, resulting in nine EMFs. The first six EMFs contained the main vibration characteristics. The technical team calculated the energy distribution of each frequency component and found that the low-frequency band (0.1–50 Hz) accounted for 42% of the total energy, the mid-frequency band (50–500 Hz) accounted for 35%, and the high-frequency band (500–3200 Hz) accounted for 23%. Combined with phase feature analysis, a multidimensional vibration feature vector containing 18 characteristic parameters was formed, as shown in Table 1.

[0089] Table 1. Parameter Table of Multidimensional Vibration Eigenvector

[0090] Based on the operating characteristics of thermal energy storage systems, the technical team established a rule for generating operating condition identification codes, including eight parameters such as load level, molten salt flow rate, temperature difference, and pump speed. Load levels are divided into five levels (0–20%, 20–40%, 40–60%, 60–80%, 80–100%), and molten salt flow rates are divided into four levels (0–500, 500–1200, 1200–2000, 2000–2800). The temperature difference is divided into three ranges (0-100, 100-200, and 200-290℃), and the pump speed is divided into six levels. 360 different operating condition identification codes are generated through combination coding.

[0091] The longest common subsequence algorithm was used to perform operating condition matching in a multi-condition vibration fingerprint database containing 28,800 sets of historical data, achieving a matching accuracy of 91.2%. The historical operating condition benchmark data includes reference indicators such as standard vibration amplitude, spectral distribution, and phase relationship for each node. The technical team found that the similarity between the current operating condition and the historical operating condition was 0.87, meeting the benchmark matching requirements.

[0092] In the two-level game optimization phase, the technical team established an upper-level game model with the goal of maximizing the accuracy of vibration anomaly detection, and the objective function included the number of actual anomaly detections. The total number of anomalies is 146. The detection threshold parameter is 162. The optimized value is 0.34, which is the standard threshold reference value. The value is set to 0.5. The lower-level game model aims to minimize the false alarm rate, and the number of false alarms is... The total number of normal samples should be kept below 23. The value was 1847. Through iterative optimization, the optimal anomaly detection threshold was determined to be 0.31, at which point the detection accuracy reached 90.1% and the false alarm rate was reduced to 1.2%.

[0093] The vibration anomaly identification model employs a Transformer encoder structure based on a multi-head attention mechanism, containing 6 encoder layers. The number of heads in the multi-head attention mechanism is set to 8, and the hidden layer dimension is 512. The model training dataset contains 9640 normal vibration samples and 2360 abnormal vibration samples, divided into training, validation, and test sets in an 8:1:1 ratio. Training is performed using the Adam optimizer with a learning rate of 0.001 and a batch size of 64. After 187 training epochs, the model converged, achieving a validation set accuracy of 92.8%.

[0094] In actual operation, the technical team inputs the multi-dimensional vibration feature vectors collected in real time into the trained vibration anomaly recognition model. For example... Figure 3 As shown, the model outputs the anomaly probability distribution for each node in the pipeline network. The anomaly probability for node N15 is 0.73, for node N22 it is 0.68, for node N08 it is 0.45, and for the remaining nodes it is below 0.3. The anomaly type prediction results indicate that node N15 exhibits pipeline resonance anomaly, and node N22 exhibits fluid impact anomaly.

[0095] Based on finite element transfer function calculations, the propagation velocity of vibration in the stainless steel pipe is 5120. The attenuation coefficient in the main pipeline is 0.023. In the branch pipe, it is 0.041. Based on the anomaly probability distribution, Dijkstra's shortest path algorithm was used to trace the location of the anomaly source, which was determined to be at the T-junction connection near node N15, approximately 2.3m from that node. On-site inspection confirmed that there was indeed a pipe vibration problem caused by loose support components at this location.

[0096] The technical team calculated that the deviation between the current anomaly probability distribution and the historical operating condition baseline was 0.69, triggering a secondary detection process based on the discrimination criteria. The secondary detection, employing a higher sampling frequency of 51200Hz and finer frequency band decomposition, confirmed the existence of the anomaly. Subsequently, the system automatically adjusted the monitoring frequency, changing from once per hour to once every 20 minutes, increasing the monitoring frequency by 200%.

[0097] After handling the anomaly, the technical team added the multidimensional vibration feature vectors from this abnormal operating condition to the multi-condition vibration fingerprint database. They then used a sliding window mechanism to delete the earliest 1200 data sets, keeping the database size below 30,000 samples. The updated database contains more anomaly patterns, improving the accuracy of subsequent anomaly identification.

[0098] During three consecutive months of operation, the vibration anomaly identification system detected 17 vibration anomaly events, of which 15 were genuine anomalies and 2 were false alarms, achieving a detection accuracy rate of 88.2%. The system successfully detected various anomalies in advance, including loose pipe supports, worn pump bearings, and internal valve leaks, with an average early warning time of 4.8 hours, providing ample time for equipment maintenance. As shown in Table 2, the detection effectiveness varied among different types of anomalies.

[0099] Table 2 Statistical Table of Detection Results for Different Anomaly Types

[0100] Compared to traditional methods of periodic manual inspection and single-point threshold alarms, this invention represents a significant technological advancement. Traditional methods rely heavily on experience and fixed threshold settings, making them susceptible to changes in operating conditions and environmental interference, and struggling to accurately identify complex vibration anomaly patterns. This invention, through multi-condition benchmark matching technology, dynamically adjusts the discrimination criteria based on actual operating conditions, avoiding misjudgments caused by fixed thresholds. A two-layer game-theoretic optimization mechanism effectively controls the false alarm rate while ensuring detection accuracy, solving the technical challenge of balancing accuracy and stability in traditional methods. Vibration propagation modeling based on graph theory and the shortest path anomaly source localization algorithm automate the entire process from anomaly detection to anomaly localization, significantly improving the efficiency and accuracy of fault diagnosis. The combination of multi-dimensional feature extraction and deep learning recognition technologies enables the system to identify complex vibration patterns that are difficult to detect using traditional methods, particularly demonstrating significant advantages in multi-frequency coupling and nonlinear vibrations.

[0101] It should be noted that the variables involved in this invention are explained in detail in Table 3 below.

[0102] Table 3. Variable Explanation Table

[0103] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying abnormal vibration in pipelines of a thermal energy storage system, characterized in that, Vibration sensors are installed at key nodes of each pipe section in the thermal energy storage system, and a pipe network topology model based on graph theory is established. The raw vibration signals from each sensor are collected and denoised using wavelet packet decomposition and adaptive spectral kurtosis selection. Empirical mode decomposition is used to extract the intrinsic mode functions of the denoised vibration signals, forming a multidimensional vibration feature vector. An operating condition identification code is established based on the current system operating parameters, and the longest common subsequence algorithm is used to find historical operating condition benchmarks in the multi-operating condition vibration fingerprint database. An upper-level game model and a lower-level game model are established, and the optimal parameters of the vibration anomaly identification model are determined through two-level game optimization. The current multidimensional vibration feature vector is input into the vibration anomaly identification model to calculate the anomaly probability distribution and anomaly type prediction results for each node in the pipe network. The anomaly source location is located based on the finite element transfer function and the shortest path algorithm. The deviation metric between the current anomaly probability distribution and the historical operating condition benchmark is calculated for anomaly judgment. The vibration monitoring frequency is adjusted according to the anomaly type and severity. The multi-operating condition vibration fingerprint database is updated.

2. The method for identifying abnormal vibration in pipelines of a thermal energy storage system according to claim 1, characterized in that, The graph-based pipeline topology model specifically abstracts the thermal energy storage system pipeline network into a graph structure, with pipe segments as edges and connection points as nodes, recording the connection relationships between pipe segments and the pipeline geometric parameters.

3. The method for identifying abnormal vibration in pipelines of a thermal energy storage system according to claim 2, characterized in that, The upper-level game model aims to maximize the vibration anomaly detection accuracy, while the lower-level game model aims to minimize the false alarm rate. The objective functions are the detection accuracy function and the false alarm rate function, which are coupled through an anomaly discrimination threshold.

4. The method for identifying abnormal vibration in pipelines of a thermal energy storage system according to claim 3, characterized in that, The aforementioned adaptive spectral kurtosis selection is specifically a method for automatically determining the optimal filtering frequency band based on the statistical characteristics of the signal. It selects the frequency range containing the most fault information by calculating the spectral kurtosis values ​​of each frequency band.

5. The method for identifying abnormal vibration in pipelines of a thermal energy storage system according to claim 4, characterized in that, The vibration anomaly identification model is specifically an anomaly detection model built based on deep learning technology. It takes a multi-dimensional vibration feature vector as input and outputs anomaly probability distribution and anomaly type prediction results.

6. The method for identifying abnormal vibration in pipelines of a thermal energy storage system according to claim 5, characterized in that, The multidimensional vibration feature vector is specifically a multidimensional array containing frequency domain, time domain, and energy characteristics of the vibration signal, used to describe the comprehensive characteristics of the vibration signal.

7. The method for identifying abnormal vibration in pipelines of a thermal energy storage system according to claim 6, characterized in that, The operating condition identification code is a character sequence generated according to a fixed coding rule based on the current load level, flow rate, and temperature operating parameters of the thermal energy storage system, and is used to identify different operating conditions.

8. The method for identifying abnormal vibration in pipelines of a thermal energy storage system according to claim 7, characterized in that, The multi-condition vibration fingerprint database is specifically a knowledge base that stores the normal vibration characteristics under different conditions, including the vibration spectrum and amplitude distribution standard patterns of each condition.

9. The method for identifying abnormal vibration in pipelines of a thermal energy storage system according to claim 8, characterized in that, The finite element transfer function specifically describes the propagation characteristics of vibration signals in the pipeline structure and is used to calculate the propagation path and attenuation coefficient of vibration in the pipeline network.

10. The method for identifying abnormal vibration in pipelines of a thermal energy storage system according to claim 9, characterized in that, The deviation metric is specifically a quantitative indicator of the difference between the current abnormal probability distribution and the historical operating condition benchmark. It is calculated using Euclidean distance. When the deviation metric ∈ [0, 0.25), it is judged to be in a normal state. When the deviation metric ∈ [0.25, 0.7), a secondary detection process is triggered. When the deviation metric ∈ [0.7, 1], the presence of vibration abnormality is confirmed.

Images