Online monitoring method for abnormal vibration of heat energy storage equipment

Abstract

The invention provides an on-line monitoring method for abnormal vibration of thermal energy storage equipment, and belongs to the technical field of thermal energy storage equipment. A multi-source sensor array comprising low-frequency and high-frequency vibration sensors is arranged at key parts of the thermal energy storage equipment, and noise interference is suppressed in a differential arrangement mode; a multi-source coupling vibration signal is separated by using an independent component analysis algorithm, time-frequency analysis is performed by combining wavelet transform to extract vibration characteristics, and a phase change vibration characteristic mechanism equation is established to calculate material physical parameter changes in a phase change process. A neural network identification model based on dynamic topology reconstruction sparse connection learning and probability graph model structured prediction is constructed, and a vibration anomaly discrimination threshold system is established to realize anomaly early warning and adaptive optimization. The technical problem that the abnormal vibration mode is difficult to accurately separate and identify by the multi-source coupling vibration signal in the phase change process of the thermal energy storage equipment is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of thermal energy storage equipment, and more specifically, relates to an online monitoring method for abnormal vibration of thermal energy storage equipment. Background Technology

[0002] As a crucial component of new energy systems, the operational stability of thermal energy storage equipment directly impacts the overall system's energy efficiency. Traditional vibration monitoring methods primarily employ single-type sensors for signal acquisition, extracting vibration characteristics through time-domain or frequency-domain analysis, and utilizing statistical analysis or simple machine learning algorithms for anomaly detection. These methods are widely used in industrial equipment condition monitoring, rotating machinery fault diagnosis, and building structural health monitoring, enabling the identification of abnormal equipment conditions to a certain extent. However, in current thermal energy storage equipment monitoring technologies, the complex physicochemical changes that occur during the solid-liquid phase transition of phase change materials lead to multiple coupled vibration sources within the equipment. Traditional single-source signal analysis methods cannot effectively handle multi-source mixed vibration signals. Existing technologies typically ignore the impact mechanism of the phase transition process on vibration characteristics and lack analytical models for changes in the physical parameters of phase change materials, resulting in low accuracy in identifying abnormal vibrations during the phase transition process. In other words, existing technologies face the technical challenge of accurately identifying abnormal vibration modes from multi-source coupled vibration signals during the phase transition process of thermal energy storage equipment. Summary of the Invention

[0003] In view of this, the present invention provides an online monitoring method for abnormal vibration of thermal energy storage equipment, which can solve the technical problem in the prior art that it is difficult to accurately identify abnormal vibration modes of multi-source coupled vibration signals during the phase change process of thermal energy storage equipment.

[0004] This invention is implemented as follows: It provides an online monitoring method for abnormal vibrations in thermal energy storage equipment. A multi-source vibration sensor array is deployed at key parts of the thermal energy storage equipment to collect multi-channel vibration and temperature signals. These signals are preprocessed and digitally filtered. A signal separation matrix is ​​established based on independent component analysis (ICA) to separate the multi-source coupled vibration signals. Wavelet transform time-frequency analysis is performed on the separated vibration signals to extract vibration feature vector sets. The influence of phase change material physical state changes on vibration characteristics is analyzed using the phase change vibration characteristic mechanism equation to establish a phase change vibration characteristic parameter set. A phase change vibration feature identification model is used to process the vibration feature vector set, temperature data, and phase change vibration characteristic parameter set to identify abnormal vibration patterns. A vibration anomaly discrimination threshold system is established to trigger an abnormal vibration early warning mechanism. The working mode of the sensor array and the signal processing algorithm parameters are adaptively adjusted according to the type and severity of the abnormal vibration.

[0005] Specifically, the arrangement of the multi-source vibration sensor array involves arranging a combination of multiple vibration sensors of different types on the thermal energy storage device according to a preset spatial distribution pattern. These sensors include low-frequency vibration sensors and high-frequency vibration sensors. A differential arrangement is used to reduce environmental noise interference. At the same time, a temperature vibration coupling sensor is set in the phase change material area to monitor the vibration status of the equipment in all directions.

[0006] The differential arrangement method refers to installing sensors of the same type in pairs, and effectively suppressing common-mode noise interference through differential signal processing. The temperature-vibration coupling sensor refers to a composite sensor that simultaneously measures temperature and vibration, used to monitor the influence of temperature changes on vibration characteristics during phase transition.

[0007] Specifically, the acquisition steps for the multi-channel vibration signal and temperature signal involve synchronous sampling using a high-precision data acquisition system. The sampling frequency is set to 2.56 times the highest frequency of the vibration signal according to the Nyquist sampling theorem. The acquired signals are then preprocessed and digitally filtered.

[0008] The independent component analysis algorithm refers to a blind source separation algorithm that decomposes multiple mixed signals into statistically independent source signals based on the principle of statistical independence. It seeks a separation matrix that maximizes the statistical independence of the separated signals. The signal separation matrix refers to the linear transformation matrix used in the independent component analysis algorithm to transform the mixed signal into an independent signal.

[0009] The wavelet transform time-frequency analysis refers to a signal analysis method that uses wavelet basis functions to decompose the signal at multiple scales and simultaneously obtain the signal's time and frequency information. The vibration feature vector set refers to the set of feature parameters representing the vibration state extracted from the vibration signal, including amplitude features, frequency features, phase features, and statistical features.

[0010] The phase transition vibration characteristic mechanism equation is used to describe the influence of changes in physical parameters of the phase transition material on its vibration characteristics during the phase transition process. The inputs include the current temperature T and the phase transition temperature. The system calculates the material density ρ and elastic modulus E, outputs a set of phase transition vibration characteristic parameters, and calculates the rate of change of material density, rate of change of elastic modulus, and rate of change of damping coefficient during the phase transition process.

[0011] The phase transition vibration characteristic parameter set refers to the set of parameters characterizing the changes in the physical properties of materials during the phase transition process, calculated by the phase transition vibration characteristic mechanism equation. These parameters include the material density change rate, the elastic modulus change rate, and the damping coefficient change rate. The material density change rate refers to the proportion of the material density change relative to the initial density during the phase transition process.

[0012] The phase change vibration feature recognition model refers to a machine learning model used to identify the vibration feature change law during the phase change process of thermal energy storage materials. It learns the mapping relationship between the phase change process and vibration features through training, and outputs the phase change vibration anomaly probability and anomaly type identifier. The value range of the phase change vibration anomaly probability is ∈ [0, 1].

[0013] The phase change vibration feature recognition model is structured as a multi-layer feedforward neural network, comprising an input layer, three hidden layers, and an output layer. The input layer receives a set of vibration feature vectors, temperature data, and a set of phase change vibration characteristic parameters. The first hidden layer adopts a sparse connection learning framework based on dynamic topology reconstruction, the second and third hidden layers adopt a fully connected structure, and the output layer adopts a structured prediction framework based on a probabilistic graphical model.

[0014] The sparse connection learning framework based on dynamic topology reconstruction monitors the activation frequency and intensity of neurons during training, calculates the contribution of each connection edge to the network output, probabilistically deletes connections when the connection contribution is lower than a set threshold, and adds new connections between high-contribution neurons according to gradient information and mutual information criteria, thereby achieving dynamic optimization of the network topology.

[0015] The structured prediction framework based on the probabilistic graphical model models the correlation between multiple output labels of abnormal vibrations as an undirected graph. Each node in the graph represents a prediction label, and the edge weights represent the conditional dependencies between labels. Probabilistic information is iteratively transmitted between graph nodes through the belief propagation algorithm, and the complex joint probability distribution is approximated as an easily computed variational distribution using variational inference techniques.

[0016] The sparse connection parameters include a connection pruning threshold, a sparsity target, and a regularization weight, which are determined based on the vibration signal complexity function, the sensor quantity ratio function, and the phase transition state weight function, respectively. The vibration signal complexity function is used to evaluate the complexity of the current vibration signal. The inputs include signal variance, spectral entropy, sample entropy, and correlation dimension, and the output is a complexity index.

[0017] The vibration anomaly discrimination threshold system refers to a set of threshold parameters established based on historical operating data and statistical analysis to determine whether the vibration state is abnormal. It includes vibration amplitude threshold, frequency offset threshold and phase change vibration anomaly probability threshold. When the vibration characteristic parameters exceed the normal operating threshold range or the phase change vibration anomaly probability ∈ (0.8, 1], the abnormal vibration early warning mechanism is triggered.

[0018] The system includes steps to adjust the sampling frequency and filtering parameters of the monitoring parameters after triggering the abnormal vibration early warning mechanism. Based on the type and severity of the abnormal vibration, the system adaptively adjusts the working mode of the sensor array and the signal processing algorithm parameters to achieve dynamic optimization of the online monitoring system and precise location of the fault source.

[0019] The normal operating threshold range refers to the statistical distribution range of various vibration characteristic parameters of the equipment under normal operating conditions. Exceeding the range indicates an anomaly. The sensor quantity ratio function is used to adjust the network sparsity based on the number of effective sensors. The inputs include the total number of sensors, the number of effective sensors, and the signal quality score, and the output is the sparsity adjustment coefficient.

[0020] This invention achieves effective separation of mixed vibration signals by establishing a multi-source vibration sensor array and an independent component analysis algorithm, thus solving the problem of mutual interference between multi-source coupled signals. By introducing the phase transition vibration characteristic mechanism equation, a quantitative relationship between changes in material physical parameters and vibration characteristics during phase transition is established, overcoming the shortcomings of traditional methods that neglect the phase transition mechanism. Employing a sparse connection learning framework based on dynamic topology reconstruction and a structured prediction mechanism using a probabilistic graphical model, an identification model for phase transition vibration characteristics is constructed, significantly improving the accuracy and reliability of anomalous vibration mode identification. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a frequency band distribution diagram of the vibration amplitude threshold in the embodiment.

[0023] Figure 3 This is a time-series variation diagram of the abnormal probability of phase transition vibration in the embodiment. Detailed Implementation

[0024] 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.

[0025] like Figure 1 The diagram shows a flowchart of an online monitoring method for abnormal vibration of a thermal energy storage device provided by the present invention. This method includes the following steps: S01. A multi-source vibration sensor array is arranged in key parts of the thermal energy storage device. The multi-source vibration sensor array includes low-frequency vibration sensors and high-frequency vibration sensors. A differential arrangement is adopted to reduce environmental noise interference. At the same time, a temperature vibration coupling sensor is set in the phase change material area. S02. Collect multi-channel vibration and temperature signals, and use a high-precision data acquisition system for synchronous sampling. The sampling frequency is set to 2.56 times the highest frequency of the vibration signal according to the Nyquist sampling theorem. The collected signals are preprocessed and digitally filtered. S03. Based on the independent component analysis algorithm, separate multi-source coupled vibration signals, establish a signal separation matrix, decompose the mixed vibration signal into independent single-source vibration components, and eliminate mutual coupling interference between signals. S04. Perform wavelet transform time-frequency analysis on the separated vibration signal to extract the time-domain and frequency-domain features of the vibration signal, establish a vibration feature vector set, and combine the temperature signal to identify the vibration feature changes during the phase transition process. S05. Analyze the influence of the physical state change of phase change material on vibration characteristics using the phase change vibration characteristic mechanism equation, calculate the material density change rate, elastic modulus change rate and damping coefficient change rate during the phase change process, and establish a set of phase change vibration characteristic parameters. S06. Using the phase change vibration feature identification model to process the vibration feature vector set, temperature data and phase change vibration characteristic parameter set, identify the abnormal vibration mode in the phase change process of thermal energy storage material, and output the abnormal probability and abnormal type identifier of phase change vibration. S07. Establish a vibration anomaly discrimination threshold system. When the vibration characteristic parameters exceed the normal operation threshold range or the probability of phase change vibration anomaly ∈ (0.8, 1], trigger the abnormal vibration early warning mechanism and adjust the sampling frequency and filtering parameters of the monitoring parameters. S08. Based on the type and severity of abnormal vibration, the working mode of the sensor array and the signal processing algorithm parameters are adaptively adjusted to achieve dynamic optimization of the online monitoring system and precise location of the fault source.

[0026] The multi-source vibration sensor array refers to a combination of multiple vibration sensors of different types arranged according to a preset spatial distribution pattern on the thermal energy storage device, used for comprehensive monitoring of the device's vibration status. The differential arrangement method refers to installing sensors of the same type in pairs, effectively suppressing common-mode noise interference through differential signal processing. The temperature-vibration coupling sensor is a composite sensor that simultaneously measures temperature and vibration, used to monitor the impact of temperature changes on vibration characteristics during phase transition.

[0027] The Independent Component Analysis (ICA) algorithm refers to a blind source separation algorithm that decomposes multiple mixed signals into statistically independent source signals based on the principle of statistical independence. Its core is to find the separation matrix that maximizes the statistical independence of the separated signals. The signal separation matrix refers to the linear transformation matrix used in the ICA algorithm to transform the mixed signal into independent signals.

[0028] The wavelet transform time-frequency analysis refers to a signal analysis method that uses wavelet basis functions to decompose a signal at multiple scales, simultaneously acquiring the signal's time and frequency information. The vibration feature vector set refers to the set of characteristic parameters representing the vibration state extracted from the vibration signal, including amplitude features, frequency features, phase features, and statistical features.

[0029] The phase transition vibration characteristic mechanism equation is used to describe the influence of changes in physical parameters of the phase transition material on its vibration characteristics during the phase transition process. The inputs include the current temperature T and the phase transition temperature. Material density And the elastic modulus E, the output is a set of phase transition vibration characteristic parameters. The mechanism equation of the phase transition vibration characteristics is expressed as follows: ,in The density variation coefficient, , , These are the reference density, reference temperature, and reference elastic modulus, respectively.

[0030] The phase transition vibration characteristic parameter set refers to the set of parameters characterizing the changes in the physical properties of materials during the phase transition process, calculated through the phase transition vibration characteristic mechanism equation. These parameters include the rate of change of material density, the rate of change of elastic modulus, and the rate of change of damping coefficient. The rate of change of material density refers to the proportion of change in material density relative to the initial density during the phase transition, and is dimensionless. The rate of change of elastic modulus refers to the proportion of change in the material's elastic modulus relative to the initial elastic modulus during the phase transition, and is dimensionless. The rate of change of damping coefficient refers to the proportion of change in the material's damping coefficient relative to the initial damping coefficient during the phase transition, and is dimensionless.

[0031] The phase change vibration feature recognition model refers to a machine learning model used to identify the changes in vibration characteristics during the phase change process of thermal energy storage materials. It learns the mapping relationship between the phase change process and vibration characteristics through training. The phase change vibration anomaly probability refers to the probability value output by the phase change vibration feature recognition model, representing the degree of abnormality in the vibration state of the current phase change process, with a value range of [0, 1]. The anomaly type identifier refers to the classification label output by the phase change vibration feature recognition model, representing the type of abnormal vibration, used to distinguish different types of vibration anomalies.

[0032] The specific structure of the phase transition vibration feature recognition model is a multi-layer feedforward neural network, including an input layer, three hidden layers, and an output layer. The input layer receives a set of vibration feature vectors, temperature data, and a set of phase transition vibration characteristic parameters. The first hidden layer adopts a sparse connection learning framework based on dynamic topology reconstruction, adaptively adding and deleting connection edges according to the statistical distribution of neuron activation patterns, evaluating connection importance through mutual information criteria and implementing probabilistic connection pruning, and introducing L2 regularization constraints to maintain the network's expressive power. The second and third hidden layers adopt a fully connected structure. The output layer adopts a structured prediction framework based on a probabilistic graphical model, modeling the output space constraint relationship as edge connections of a graph, propagating probabilistic information between graph nodes through a belief propagation algorithm, and using variational inference techniques to approximate the posterior distribution.

[0033] The steps for establishing the training dataset of the phase change vibration feature recognition model specifically include collecting vibration and temperature signals of thermal energy storage equipment under different operating conditions, marking normal and abnormal operating states, extracting vibration feature vector sets and temperature feature vectors as input samples, calculating phase change vibration characteristic parameter sets using the phase change vibration characteristic mechanism equation as supplementary input, establishing sample labels corresponding to abnormal type identification and abnormal severity, and dividing the dataset into training set, validation set and test set in a ratio of 7:2:1.

[0034] The specific steps for training the phase transition vibration feature recognition model include initializing network parameters, setting the learning rate to 0.001, the batch size to 64, using the Adam optimization algorithm for gradient descent, and using a weighted combination of cross-entropy loss and structured loss as the loss function. During training, the sparse connection parameters and structured prediction parameters are adjusted based on the performance of the validation set. After 500 training iterations, the optimal model parameters are saved.

[0035] The sparse connection learning framework based on dynamic topology reconstruction monitors the activation frequency and intensity of neurons during training, calculates the contribution of each connection edge to the network output, and probabilistically deletes connections when their contribution falls below a set threshold. Simultaneously, new connections are added between high-contribution neurons based on gradient information and mutual information criteria, thus dynamically optimizing the network topology. The connection pruning probability is jointly determined by the connection importance score and the global sparsity objective. Grouping regularization constraints are introduced to ensure that each neuron maintains at least a certain number of effective connections, preventing network degradation. This mechanism can adaptively adjust network complexity, significantly reducing computational overhead while maintaining prediction accuracy, making it particularly suitable for processing high-dimensional vibrational feature data.

[0036] The structured prediction framework based on a probabilistic graphical model models the correlation between multiple output labels of anomalous vibrations as an undirected graph. Each node in the graph represents a predicted label, and the edge weights represent the conditional dependencies between labels. During the prediction phase, a belief propagation algorithm iteratively transmits probabilistic information between graph nodes. Variational inference techniques are used to approximate the complex joint probability distribution into an easily computed variational distribution, and a mean-field approximation assumption simplifies the inference process. This mechanism ensures the logical consistency of the prediction results, avoids contradictory results from independent predictions, and significantly improves the accuracy and reliability of anomalous vibration type identification. The synergistic effect of these two mechanisms enables the phase change vibration feature identification model to possess both efficient feature learning capabilities and the ability to output structured prediction results, effectively solving the challenges of multi-source signal coupling and fault mode identification in vibration monitoring of thermal energy storage equipment.

[0037] The sparse connection parameters include a connection pruning threshold, a sparsity target, and regularization weights, which are determined based on the vibration signal complexity function, the sensor number ratio function, and the phase transition state weight function, respectively. The vibration signal complexity function is used to evaluate the complexity of the current vibration signal. Its inputs include signal variance, spectral entropy, sample entropy, and correlation dimension, and its output is a complexity index, ranging from [0, 1]. The sensor number ratio function is used to adjust the network sparsity based on the number of effective sensors. Its inputs include the total number of sensors, the number of effective sensors, and the signal quality score, and its output is a sparsity adjustment coefficient, ranging from (0, 2). The phase transition state weight function is used to adjust the model parameters based on the phase transition process. Its inputs include the current temperature T and the phase transition temperature. The phase transition process and phase transition rate are output as parameter weighting coefficients, with values ​​ranging from [0.5, 1.5].

[0038] The vibration anomaly discrimination threshold system refers to a set of threshold parameters established based on historical operating data and statistical analysis to determine whether the vibration state is abnormal. This includes vibration amplitude thresholds, frequency offset thresholds, and phase change vibration anomaly probability thresholds. The normal operating threshold range refers to the statistical distribution interval of each vibration characteristic parameter of the equipment under normal operating conditions; values ​​exceeding this interval indicate a possible anomaly.

[0039] In addition, the present invention also provides a computer-based method for forming an online monitoring system for abnormal vibration of thermal energy storage equipment. 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-described method.

[0040] The specific implementation methods of the above steps are described in detail below.

[0041] The specific implementation of step S01 involves installing a multi-source vibration sensor array on key components of the thermal energy storage device, such as the shell surface, internal support structure, and phase change material container wall. First, the sensor placement is determined based on the device's structural characteristics and vibration propagation path analysis. Low-frequency vibration sensors, with a frequency response range of 0.1Hz to 1000Hz, are arranged on the device surface according to spatial geometric symmetry principles to monitor the overall low-frequency vibration modes of the device. Simultaneously, high-frequency vibration sensors, with a frequency response range of 1000Hz to 50000Hz, are installed in highly vibration-sensitive areas to capture localized high-frequency vibration anomalies. When using a differential arrangement, sensors of the same type are installed in pairs in adjacent positions, with the distance between the two sensors controlled within 5% to 10% of the device's characteristic dimensions. Differential signal processing technology is used to eliminate common-mode noise such as environmental vibration and electromagnetic interference. A temperature-vibration coupling sensor is installed in the phase change material region. This sensor integrates a thermocouple temperature measurement element and a piezoelectric vibration measurement element, enabling simultaneous monitoring of temperature changes during the phase change process and local vibration changes caused by the phase change. The spatial distribution of the sensor array follows the Nyquist spatial sampling theorem to ensure that the spatial sampling frequency meets the vibration wavelength resolution requirements, and the sensor spacing does not exceed half of the shortest wavelength of interest.

[0042] The specific implementation of step S02 involves using a high-precision multi-channel data acquisition system to digitally acquire the analog signals output by the vibration and temperature sensors. The data acquisition system employs a 24-bit analog-to-digital converter with a dynamic range of 144 dB and an input impedance greater than 1 megaohm, ensuring high fidelity signal acquisition. Based on the Nyquist sampling theorem, the sampling frequency is set to 2.56 times the highest frequency of the vibration signal; for a high-frequency vibration signal with a highest frequency of 50,000 Hz, the sampling frequency is set to 128,000 Hz. Hardware clock synchronization technology is used between channels to achieve synchronous sampling, with the time deviation between channels controlled within 1 microsecond. Signal preprocessing includes DC bias removal, anti-aliasing filtering, and amplitude calibration. The anti-aliasing filter uses an 8th-order elliptic filter with a passband ripple of less than 0.1 dB and a stopband attenuation greater than 80 dB. The digital filter adopts a finite impulse response filter design. It applies a bandpass filter of 0.1Hz to 1000Hz to low-frequency vibration signals and a bandpass filter of 1000Hz to 25000Hz to high-frequency vibration signals. The transition bandwidth of the filter is set to 10% of the passband width.

[0043] The specific implementation of step S03 involves using the Independent Component Analysis (ICA) algorithm to perform blind source separation processing on the multi-channel vibration signal. Based on the principle of statistical independence, the ICA algorithm finds the optimal separation matrix by maximizing the non-Gaussianity of the separated signals. The algorithm first centers the multi-channel mixed signal, removing the mean of each channel signal, and then performs whitening preprocessing to transform the signal covariance matrix into an identity matrix, reducing the second-order statistical correlation between signals. The separation matrix is ​​solved using a fast fixed-point algorithm, which iteratively optimizes the objective function to find the transformation matrix that maximizes the statistical independence of the separated signals. The objective function is chosen based on a non-Gaussianity metric of negative entropy, and its convergence criterion is that the change in the separation matrix between two consecutive iterations is less than... The maximum number of iterations during the separation process was set to 1000, and the learning rate was adaptively adjusted from 0.001 to 0.1. The dimension of the signal separation matrix was equal to the number of sensor channels. This matrix decomposed the mixed vibration signal into statistically independent single-source vibration components, effectively eliminating coupling interference between different vibration sources. The separation effect was evaluated using statistical indicators such as mutual information and kurtosis. A mutual information value less than 0.05 indicated a good separation effect.

[0044] The specific implementation of step S04 involves performing wavelet transform time-frequency analysis on the vibration signal after independent component analysis to extract characteristic parameters representing the vibration state. The wavelet transform uses the Morlet wavelet as the mother wavelet function, which has good time-frequency resolution characteristics and is suitable for analyzing non-stationary vibration signals. The scale parameter range of the wavelet transform is set to 1 to 1024, corresponding to a frequency range of 0.1Hz to 50000Hz, with the time resolution adaptively adjusted according to the scale parameter. Time-domain feature extraction includes statistical features such as the signal's effective value, peak value, kurtosis, and skewness. The effective value reflects the vibration energy level, the peak value represents the maximum vibration amplitude, and kurtosis and skewness describe the probability distribution characteristics of the vibration signal. Frequency-domain feature extraction includes parameters such as dominant frequency, spectral centroid, and spectral entropy. The dominant frequency is obtained through peak detection of the power spectral density function, the spectral centroid is calculated as the power-weighted average of each frequency component, and the spectral entropy reflects the complexity of the frequency components. Simultaneously, combined with the temperature signal from the temperature-vibration coupling sensor, the vibration characteristic changes caused by material property changes during the phase transition are identified. The start and end times of the phase transition are detected by temperature gradient detection, and the differences in vibration characteristics before and after the phase transition are analyzed. The established vibration feature vector set contains 32 feature parameters, covering a comprehensive feature description in the time domain, frequency domain, and time-frequency domain.

[0045] The specific implementation of step S05 is based on a theoretical model of the physical property changes of phase change materials, calculating the influence of material parameter changes on vibration characteristics during the phase change process. The phase change vibration characteristic mechanism equation describes the changes in physical parameters such as material density, elastic modulus, and damping coefficient during the phase change process. Input parameters include the current temperature T and the phase change temperature. Initial material density Initial elastic modulus Physical parameters. The calculation of the material density change rate considers the volume change caused by the solid-liquid transition during the phase transition, and the density change coefficient. The typical value range is 0.05 to 0.15. The rate of change of the elastic modulus reflects the change in material stiffness during the phase transition process; the elastic modulus of liquid materials is typically 2 to 3 orders of magnitude lower than that of solid materials. The rate of change of the damping coefficient describes the change in the material's internal friction characteristics; the damping coefficient increases due to intensified molecular motion during the phase transition. Reference temperature. Set to ambient temperature 25 degrees Celsius, reference density and reference elastic modulus The vibration characteristic parameters are determined based on the physical properties of the phase change material. The set of phase change vibration characteristic parameters calculated by this mechanism equation provides a physical basis for subsequent vibration characteristic identification and establishes a quantitative relationship between vibration characteristics and changes in material properties.

[0046] The specific implementation of step S06 involves using a phase transition vibration feature recognition model to process the vibration feature vector set, temperature data, and phase transition vibration characteristic parameter set calculated in step S05, obtained in step S04, to identify abnormal vibration modes during the phase transition process of thermal energy storage materials. This recognition model employs a multi-layer feedforward neural network structure, capable of learning complex nonlinear mapping relationships. The model takes the vibration feature vector set, normalized temperature data, and phase transition vibration characteristic parameter set as input, and integrates different types of input data into a unified feature representation through a feature fusion layer. The first hidden layer of the network uses a sparse connection learning framework based on dynamic topology reconstruction, dynamically adjusting connection weights according to the activation patterns of neurons during training, evaluating connection importance through mutual information criteria, and implementing probabilistic connection pruning. The second and third hidden layers employ a fully connected structure, using a modified linear unit activation function, with each layer containing 128 neurons. The output layer adopts a structured prediction framework based on a probabilistic graphical model, modeling the logical relationships between multiple output labels as a graph structure, and ensuring the consistency of prediction results through a belief propagation algorithm. The model output includes the probability of phase transformation vibration anomaly and the anomaly type identifier. The anomaly probability ranges from 0 to 1, and the anomaly type identifier includes categories such as mechanical loosening, material fatigue, and phase transformation anomaly.

[0047] The specific implementation of step S07 involves establishing a multi-level vibration anomaly discrimination threshold system, making a comprehensive judgment based on vibration characteristic parameters and the probability of phase transition vibration anomalies. The vibration amplitude threshold is determined based on the vibration statistical characteristics during normal equipment operation, using a 3-standard-deviation criterion to set the warning threshold. An early warning is triggered when the effective vibration value exceeds the mean plus 3 standard deviations. The frequency offset threshold is set to 10% of the main frequency deviating from the normal value; a frequency anomaly is judged when the detected main frequency offset exceeds this threshold. The phase transition vibration anomaly probability threshold is set to 0.8; an abnormal vibration early warning mechanism is triggered when the model outputs an anomaly probability greater than 0.8. The early warning mechanism includes measures such as increasing the sampling frequency, adjusting filter parameters, and increasing sensor monitoring frequency. The sampling frequency is adaptively adjusted according to the severity of the anomaly; for minor anomalies, the sampling frequency is increased to 1.5 times the original, and for severe anomalies, it is increased to 2 times. Filter parameter adjustments include narrowing the filter passband range and improving filter accuracy to more accurately capture abnormal vibration characteristics. Simultaneously, information such as the anomaly occurrence time, anomaly type, and anomaly severity is recorded to provide a basis for equipment maintenance decisions.

[0048] The specific implementation of step S08 involves adaptively adjusting the operating mode and algorithm parameters of the entire monitoring system based on the identified abnormal vibration type and severity. For mechanical loosening anomalies, the weight of low-frequency vibration sensors is increased to improve monitoring accuracy in the 0.1Hz to 100Hz frequency band. For material fatigue anomalies, high-frequency vibration components are emphasized, and the wavelet transform scaling parameters are adjusted to improve high-frequency resolution. For phase transition anomalies, the data acquisition frequency of the temperature-vibration coupling sensor is increased, and the temperature sampling interval is shortened to 1 second. The sensor array operating modes include normal mode, early warning mode, and emergency mode. In normal mode, the sensors operate at the standard sampling frequency; in early warning mode, the sampling frequency is increased by 25%; and in emergency mode, the sampling frequency is increased by 50% and all backup sensors are activated. The adaptive adjustment of signal processing algorithm parameters includes learning rate adjustment for independent component analysis, optimization of wavelet transform scaling parameters, and threshold updates for neural network models. The learning rate is adaptively adjusted according to signal complexity; the learning rate is reduced to improve stability when complexity is high, and increased to accelerate convergence when complexity is low. Accurate fault location is achieved by analyzing the intensity and time difference of abnormal signals detected by each sensor. The spatial coordinates of the abnormal source are calculated using a triangulation algorithm, and the positioning accuracy can reach within 5% of the equipment's feature size.

[0049] Further explanation is needed regarding the detailed structure of the phase transition vibration feature recognition model, which comprises a multi-layer feedforward neural network architecture consisting of an input layer, three hidden layers, and an output layer. The input layer receives three types of data: a 32-dimensional vibration feature vector set, a temperature data sequence, and a 6-dimensional phase transition vibration characteristic parameter set. After feature normalization and data fusion, a 64-dimensional comprehensive input vector is formed. The first hidden layer employs a sparse connection learning framework based on dynamic topology reconstruction, containing 256 neurons. The connection weights are dynamically adjusted using a mutual information criterion. This layer monitors the activation frequency and intensity of each neuron during training, calculates the contribution of each connection edge to the network output, and probabilistically deletes connections when the contribution is below 0.01. Simultaneously, new connections are added between high-contribution neurons based on gradient information. The connection pruning probability is determined by both the connection importance score and the global sparsity target, which is set to 0.3, maintaining a connection density of 30%. The second and third hidden layers both use a fully connected structure, containing 128 and 64 neurons respectively. Modified linear unit activation functions and batch normalization techniques are used to improve training stability. The output layer employs a structured prediction framework based on a probabilistic graphical model, comprising eight output nodes corresponding to anomaly probabilities and seven anomaly type identifiers. This layer models the logical constraints between output labels as an undirected graph, transmits probabilistic information between graph nodes through a belief propagation algorithm, and uses variational inference techniques to approximate the posterior distribution, ensuring the logical consistency of the prediction results.

[0050] The detailed steps for establishing the training dataset first involve collecting operational data of the thermal energy storage equipment under different operating conditions. These conditions include six typical operating states: normal heat charging, normal heat dissipation, variable load, start-up and shutdown, ambient temperature change, and partial load. Vibration and temperature signals were collected continuously for 72 hours under each condition, with a sampling frequency of 128,000 Hz, resulting in a total of 432 hours of raw data. The equipment operating states were manually labeled, categorizing the data into two main types: normal operation and abnormal operation. Abnormal operation was further subdivided into seven types: mechanical loosening, bearing failure, pipeline leakage, phase transition anomaly, resonance anomaly, motor failure, and seal failure. Each anomaly type was obtained through fault injection methods, including artificially introducing mechanical loosening, simulating bearing wear, and creating micro-leakage points. Using the method in step S04, a 32-dimensional vibration feature vector set was extracted from the raw signals, along with a temperature feature vector including five parameters: mean temperature, temperature gradient, and temperature change rate. The 6-dimensional phase transition vibration characteristic parameter set for each moment was calculated using the phase transition vibration characteristic mechanism equation as supplementary input features. The sample labels include binary normal / abnormal identifiers, multi-class abnormality type identifiers, and continuous value abnormality severity scores. The abnormality severity score uses a continuous value from 0 to 1, where 0 represents a normal state and 1 represents a severe abnormality. A dataset of 100,000 samples is constructed and divided into training, validation, and test sets in a 7:2:1 ratio to ensure a relatively consistent distribution of various operating conditions and abnormality types across the three datasets. The training set is used for model parameter learning, the validation set for hyperparameter tuning and model selection, and the test set for final performance evaluation.

[0051] It should be noted that the key technical ideas of this invention include a differential arrangement strategy for multi-source vibration sensor arrays, multi-source signal separation technology based on independent component analysis, a modeling method for phase transition vibration characteristics, and a neural network architecture for dynamic topology reconstruction. The differential arrangement strategy for multi-source vibration sensor arrays, by installing sensors of the same type in pairs and employing differential signal processing technology, effectively suppresses common-mode interference such as environmental noise and electromagnetic interference. Compared with traditional single-point monitoring methods, it significantly improves signal quality and monitoring accuracy, solving the problem of vibration signals being easily interfered with in the complex electromagnetic environment of thermal energy storage equipment. The multi-source signal separation technology based on independent component analysis can decompose the mixed vibration signals collected by multiple sensors into statistically independent single-source vibration components. Compared with traditional frequency domain filtering methods, it can better handle vibration sources with overlapping spectra, effectively eliminating coupling interference between different vibration sources and improving the accuracy and reliability of fault feature extraction. The modeling method for phase transition vibration characteristics establishes a quantitative relationship between changes in material properties and changes in vibration characteristics, combining physical mechanisms with data-driven methods. Compared with purely data-driven methods, it has stronger interpretability and generalization ability, accurately identifying vibration changes caused by phase transition processes and avoiding misjudging normal phase transition processes as equipment failures. The neural network architecture with dynamic topology reconstruction adaptively adjusts the network connection structure and parameters, maintaining high-precision feature learning capabilities while significantly reducing computational complexity. Compared with fixed-structure neural networks, it has better efficiency and adaptability when processing high-dimensional vibration features.

[0052] The synergistic effect of these four key technological approaches forms a complete vibration monitoring technology system for thermal energy storage equipment, exhibiting significant comprehensive advantages compared to existing technologies. The differential arrangement strategy provides high-quality raw signals, independent component analysis technology achieves effective signal separation, mechanistic modeling methods provide physical constraints and prior knowledge, and the dynamic neural network architecture enables intelligent recognition. These four elements work together to form a complete technological chain from signal acquisition to intelligent recognition. This synergistic technology system not only addresses the shortcomings of traditional methods in multi-source coupled signal processing but also improves the reliability and accuracy of monitoring results through the deep integration of physical mechanisms and artificial intelligence. It provides strong technical support for the safe and stable operation of thermal energy storage equipment and has greater practicality and promotional value compared to existing single-technology methods.

[0053] It should be noted that this invention also solves the following technical problems: First, it addresses the issue of unstable signal quality in traditional vibration monitoring systems under complex and variable environments. By employing a multi-source vibration sensor array with a differential arrangement, common-mode noise interference is effectively suppressed, improving the stability and reliability of signal acquisition. Simultaneously, the introduction of a temperature-vibration coupling sensor enables synchronous acquisition of vibration and temperature signals, providing multi-dimensional data support for phase change process monitoring. This sensor arrangement strategy, combined with the synchronous sampling mechanism of a high-precision data acquisition system, ensures high-quality monitoring data is obtained even under complex operating environments in thermal energy storage equipment. Second, it solves the problem of existing anomaly detection methods being unable to adapt to dynamic changes in equipment operating conditions. By establishing an adaptive parameter adjustment mechanism based on a vibration signal complexity function, a sensor quantity ratio function, and a phase change state weighting function, dynamic optimization of monitoring system parameters is achieved. When abnormal vibration is detected, the system can adaptively adjust the sensor array operating mode and signal processing algorithm parameters according to the anomaly type and severity, improving monitoring accuracy and response speed. This adaptive mechanism enables the monitoring system to maintain optimal monitoring performance under different operating conditions, significantly improving the system's robustness and practicality.

[0054] Specifically, the principle of this invention is as follows: This invention solves the technical problem of accurately separating and identifying abnormal vibration modes of multi-source coupled vibration signals during the phase change process of thermal energy storage equipment. Its fundamental principle lies in constructing a complete multi-source signal processing and phase change mechanism modeling system. First, a differentially arranged multi-source vibration sensor array achieves omnidirectional signal acquisition, effectively suppressing environmental noise interference and providing a high-quality data source for subsequent signal processing. Second, the independent component analysis algorithm decomposes the mixed vibration signal into independent single-source components based on the principle of statistical independence, eliminating mutual coupling interference between signals and solving the key problem of difficult separation of multi-source signals. The establishment of the phase change vibration characteristic mechanism equation is the core innovation of this invention. This equation quantitatively describes the influence of changes in physical parameters such as material density and elastic modulus on vibration characteristics during the phase change process, providing physical mechanism support for vibration signal analysis. Combined with real-time monitoring data from temperature-vibration coupled sensors, it can accurately identify vibration characteristic change modes during the phase change process, making up for the lack of phase change mechanism analysis in traditional methods. The phase transition vibration feature recognition model employs a multi-layer feedforward neural network structure. An adaptive optimization of the network structure is achieved through a sparse connection learning framework based on dynamic topology reconstruction. The network topology is dynamically adjusted according to the complexity of the vibration signal, improving feature learning efficiency. The structured prediction framework based on a probabilistic graphical model models the correlation between multiple output labels as a graph structure. A belief propagation algorithm ensures the logical consistency of the prediction results, avoiding contradictory results from independent predictions. This synergistic effect of the dual mechanisms enables the model to possess both efficient feature extraction capabilities and output structured prediction results, effectively solving the problem of accurate identification of complex vibration modes during phase transitions.

[0055] 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.

[0056] In this embodiment, the specific implementation of step S01 is the same as described above, and will not be repeated in detail here.

[0057] The specific implementation of step S02 involves using a high-precision multi-channel data acquisition system for synchronous sampling. The sampling frequency is determined according to the Nyquist sampling theorem, and the formula is expressed as follows: ; In the formula, Sampling frequency, in Hz; The highest frequency of the vibration signal, measured in Hz; This is the sampling multiple, dimensionless, with an empirical value of 2.56.

[0058] The specific implementation of step S03 is to use the independent component analysis algorithm to separate multi-source signals. The formula for calculating the signal separation matrix is ​​as follows: ; In the formula, The separated independent signal matrix has the following dimensions: same; It is a separation matrix, dimensionless; This is a mixed signal matrix, with units of... The separation matrix is ​​solved by maximizing the non-Gaussianity objective function, which is expressed as follows: ; In the formula, The objective function is dimensionless. For separate vectors, the dimensions are... The reciprocals are the same; The input signal vector, in units of ; It is a nonlinear function, usually taken as Dimensionless; A standard Gaussian random variable, dimensionless; For expectation operators.

[0059] The specific implementation of step S04 involves performing wavelet transform time-frequency analysis on the separated signals. The wavelet transform formula is expressed as follows: ; In the formula, The continuous wavelet transform coefficients have dimensions similar to the input signal. same; This is a scale parameter, dimensionless; This is a time-shift parameter, in seconds. This is a time-domain signal, with units of . ; It is the complex conjugate of the mother wavelet, dimensionless; The variable is time, and the unit is seconds (s). In this formula... The term is used to preserve the energy conservation property of wavelet transform. The formula for constructing the vibration eigenvector is expressed as follows: ; In the formula, These are normalized eigenvectors, dimensionless; For the first Each feature parameter has a dimension that depends on the feature type. For the first Reference values ​​for each characteristic parameter, with units of... same; This represents the total number of feature parameters, dimensionless, with a default value of 32.

[0060] The specific implementation of step S05 is to calculate the material property changes using the phase transition vibration characteristic mechanism equation. The phase transition vibration characteristic mechanism equation is expressed as follows: ; In the formula, This is the change in density, in units of... ; For reference density, the unit is... ; The current temperature is in Kelvin (K). This is the phase transition temperature, expressed in Kelvin (K). This is a reference temperature, in K, typically taken as 298K; This is the density variation coefficient, dimensionless, with an empirical value of 0.05–0.15; This is the current elastic modulus, in Pa. The elastic modulus is used as a reference, with units of Pa. The formula for calculating the rate of change of material density is as follows: ; In the formula, This is the dimensionless rate of change of density; The maximum density change rate is dimensionless and has an empirical value of 0.2. The formula for calculating the rate of change of the elastic modulus is as follows: ; In the formula, The dimensionless rate of change of elastic modulus; This represents the maximum change in elastic modulus, expressed in Pa. The formula for calculating the rate of change of the damping coefficient is as follows: ; In the formula, The rate of change of the dimensionless damping coefficient; The current damping coefficient is dimensionless. The reference damping coefficient is dimensionless and has an empirical value of 0.02. This represents the maximum change in damping coefficient, which is dimensionless and has an empirical value of 0.05.

[0061] The specific implementation of step S06 involves employing a phase transition vibration feature recognition model. Within this model, the sparse connection learning framework based on dynamic topology reconstruction includes calculating the statistical distribution of neuron activation, evaluating connection importance, and probabilistic connection pruning. The formula for calculating the neuron activation frequency is as follows: ; In the formula, For the first Layer The activation frequency of a single neuron, dimensionless; The number of samples in the batch is dimensionless. It is an indicator function, dimensionless; For the first The sample at the th Layer The activation value of a single neuron is dimensionless. The activation threshold is dimensionless and has an empirical value of 0.1. The formula for calculating neuron activation intensity is as follows: ; In the formula, For the first Layer The activation intensity of each neuron is dimensionless. Connection importance is assessed based on the mutual information criterion, and the calculation formula is as follows: ; In the formula, For connection Mutual information, dimensionless; It is an empirical joint probability distribution, dimensionless; and Let be an empirical marginal probability distribution, dimensionless. The probability calculation formula for probabilistic join pruning is expressed as follows: ; In the formula, For connection The pruning probability is dimensionless. It is the sigmoid function, which is dimensionless; This is the importance weighting coefficient, dimensionless, with an empirical value of 2.0; is the sparsity weighting coefficient, dimensionless, with an empirical value of 1.5; The target sparsity is dimensionless and defaults to 0.3. The loss function for grouped regularization constraints is expressed as follows: ; In the formula, The loss is for grouped regularization and is dimensionless. This is the regularization weight, dimensionless, with an empirical value of 0.01; The total number of groups of neurons, dimensionless; For the first Group of neurons; The connection weights are dimensionless.

[0062] The specific implementation of the structured prediction framework based on probabilistic graphical models includes graph structure modeling, message passing algorithms, and variational inference computation. The formula for constructing the adjacency matrix of a graph is expressed as follows: ; In the formula, It is a dimensionless graph adjacency matrix; Output tags and The correlation coefficient between them is dimensionless; The edge connection threshold is dimensionless and has an empirical value of 0.3. The number of output labels is dimensionless and set to 8. The iterative update formula for the message passing algorithm is expressed as follows: ; In the formula, For the first In the next iteration, from node To the node The message being transmitted is dimensionless. For nodes The local potential function is dimensionless; For the edge The paired potential function is dimensionless; For nodes The set of neighboring nodes; For nodes The state variables are dimensionless. The objective function for variational inference is expressed as follows: ; In the formula, It is the optimal variational distribution, dimensionless; It is a family of variational distributions; The Kullback-Leibler divergence is dimensionless. For a given observation The posterior distribution of is dimensionless; These are latent variables, dimensionless. The formula for calculating the conditional probability of Gibbs sampling is as follows: ; In the formula, For the first During the second sampling The values ​​of the variables are dimensionless. For the first During the second sampling, except for the first All variable values ​​other than the individual variable are dimensionless. These are model parameters, dimensionless.

[0063] The specific implementation of step S07 is to establish a vibration anomaly discrimination threshold system. The formula for calculating the vibration amplitude threshold is as follows: ; In the formula, This is the dimensionless vibration amplitude threshold. The average vibration amplitude is expressed in units of 1000 m / s. ; The standard deviation of vibration amplitude is expressed in units of 1000 m / s. ; For reference vibration amplitude, the unit is... The formula for calculating the frequency offset threshold is as follows: ; In the formula, This is the dimensionless frequency offset threshold, set to 0.1. This is the current clock speed, measured in Hz. This is the normal clock speed, measured in Hz.

[0064] The specific implementation of step S08 is to adaptively adjust the system parameters according to the anomaly type. The formula for adjusting the sampling frequency is expressed as follows: ; In the formula, The adjusted sampling frequency is expressed in Hz. The initial sampling frequency is expressed in Hz. The adjustment coefficient is dimensionless and has an empirical value of 0.5 to 1.0. The degree of abnormality is dimensionless and ranges from 0 to 1.

[0065] The specific implementation of the vibration signal complexity function comprehensively considers the statistical and spectral characteristics of the signal, and the calculation formula is expressed as follows: ; In the formula, It is a dimensionless complexity exponent, with a value range of 0 to 1; These are weighting coefficients, dimensionless, with a default value of 0.25. The signal variance is expressed in units of 1 / 2 Ω. ; The variance of the reference signal is expressed in units of . ; The spectral entropy is dimensionless. The reference spectral entropy is dimensionless and has an empirical value of 0.8. The sample entropy is dimensionless. The entropy of the reference sample is dimensionless and has an empirical value of 1.2. The relevant dimension is dimensionless; For reference, the dimension is dimensionless, and the empirical value is 2.5.

[0066] The specific implementation of the sensor quantity ratio function is to adjust the network sparsity based on the proportion of effective sensors. The calculation formula is as follows: ; In the formula, This is a dimensionless sparsity adjustment coefficient, with a value ranging from 0 to 2; The number of effective sensors is expressed in units. This represents the total number of sensors, expressed in units. The reference sensor efficiency is dimensionless and has an empirical value of 0.9. The signal quality score is dimensionless. The reference signal quality score is dimensionless and has an empirical value of 0.8.

[0067] The specific implementation of the phase transition state weighting function involves dynamically adjusting the model parameters according to the phase transition process. The calculation formula is expressed as follows: ; In the formula, This is a dimensionless parameter weighting coefficient, with a value range of 0.5 to 1.5; This represents the phase transition temperature range, expressed in K, with an empirical value of 10K. The phase transition rate is expressed in K / s. The reference phase transition rate is expressed in K / s, with an empirical value of 0.1 K / s.

[0068] It should be explained that the phase transition vibration mechanism equation is established based on the theoretical analysis of the changes in physical parameters of phase transition materials during the phase transition process. This equation considers the nonlinear effects of temperature deviation from the phase transition point and the coupling effect of changes in elastic modulus. Dimensional influences are eliminated through dimensionless processing, making the equation universal. The density change term in the equation reflects the influence of material volume changes on vibration characteristics during the solid-liquid phase transition, and the temperature normalization term adopts the Lorentz function form. Describing the nonlinear characteristics of phase transition processes, elastic modulus coupling term The influence of material stiffness variation on vibration propagation is considered. Compared with the traditional linear model, this equation can accurately describe the complex physical phenomena during phase transition, providing a reliable physical basis for vibration anomaly identification.

[0069] The formulas for the rate of change of density and elastic modulus, through normalization, achieve a unified quantitative representation of the changes in different physical parameters. The formula for the rate of change of damping coefficient... This formula reflects the changes in the material's internal friction characteristics during phase transition. Compared to the traditional fixed damping model, this formula can accurately capture the dynamic changes in damping characteristics during phase transition, thus improving the accuracy of vibration characteristic analysis.

[0070] The sparse connection learning framework based on dynamic topology reconstruction utilizes the statistical distribution of neuron activation. and activation intensity This framework quantifies the activation patterns of neurons using the mutual information criterion. Assess connection importance through probabilistic pruning mechanisms Adaptive adjustment of network topology, grouping regularization constraints It ensures the network's expressive power and significantly improves computational efficiency and generalization performance compared to fixed-topology neural networks.

[0071] The structured prediction framework based on probabilistic graphical models utilizes adjacency matrices. Model the constraint relationships between output labels, and use message passing algorithms. Variational inference techniques for propagating probabilistic information between graph nodes Approximate calculation of the posterior distribution, Gibbs sampling This framework generates structured prediction results, which, compared to independent prediction methods, ensures logical consistency of the output results and significantly improves the accuracy and reliability of multi-label classification.

[0072] The wavelet transform formula achieves time-frequency domain analysis of signals through multi-scale decomposition. The choice of the mother wavelet function directly affects the time-frequency resolution. This formula, employing a continuous wavelet transform, provides more accurate time-frequency information. Compared to the traditional Fourier transform, the wavelet transform possesses time-frequency localization characteristics, effectively identifying transient features in non-stationary vibration signals, and is particularly suitable for analyzing vibration changes during phase transitions. The vibration feature vector construction formula eliminates differences in the dimensions and numerical ranges of different feature parameters through normalization processing, improving the effectiveness of feature fusion.

[0073] The objective function of Independent Component Analysis (ICA) is based on the principle of minimizing mutual information in information theory. By maximizing non-Gaussianity, it achieves blind source separation, and the algorithm can effectively separate statistically independent vibration source signals. Compared with traditional Principal Component Analysis (PCA), ICA not only considers the second-order statistical properties of the signal but also utilizes higher-order statistical information, enabling it to better handle non-Gaussian distributed vibration signals and significantly improving the accuracy of multi-source signal separation.

[0074] The vibration signal complexity function comprehensively considers both the time-domain statistical characteristics and the frequency-domain complexity characteristics of the signal, and achieves a quantitative assessment of the signal complexity by weighted fusion of multiple complexity indices. Compared with a single complexity index, this function has stronger robustness and discriminative ability, and can accurately identify different types of vibration anomaly patterns, providing reliable input features for subsequent intelligent recognition algorithms.

[0075] The calculation formulas for vibration amplitude threshold and frequency offset threshold are based on the 3σ criterion and relative deviation analysis in statistics, enabling adaptive determination of anomaly detection thresholds based on historical data. Compared to fixed threshold methods, these adaptive thresholds better adapt to changes in equipment operating conditions, reducing false alarms and false negatives, and improving the accuracy and reliability of anomaly detection.

[0076] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: A thermal energy storage system uses a mixed molten salt of sodium nitrate and potassium nitrate as the heat storage medium, with a normal operating temperature range of 290-565℃ and a phase change temperature of... The temperature is 308℃. The system mainly includes key equipment such as molten salt storage tanks, molten salt pumps, heat exchangers, and piping systems.

[0077] Following the method of this invention, the technical team first deployed a multi-source vibration sensor array at 12 key locations within the thermal energy storage device. Four low-frequency vibration sensors, operating in the range of 0.1-100Hz, were installed on the wall of the molten salt tank to monitor low-frequency structural vibrations. Six high-frequency vibration sensors, operating in the range of 10-10000Hz, were installed near the molten salt pump and heat exchanger to monitor high-frequency mechanical vibrations. A differential arrangement was used, with identical sensors installed in pairs at a spacing of 1.5m. Differential signal processing effectively reduced environmental noise interference by approximately 35dB. Two temperature-vibration coupling sensors were installed in the molten salt phase change region, capable of simultaneously measuring temperature changes and vibration signals, with a temperature measurement accuracy of ±0.5℃ and a vibration measurement accuracy of 0.001℃. .

[0078] The data acquisition system employs a 24-bit high-precision ADC. Based on the Nyquist sampling theorem, the sampling frequency is set to 25600Hz, which is 2.56 times the highest vibration signal frequency of 10000Hz. The system simultaneously acquires vibration signals from 14 channels and temperature signals from 2 channels, generating approximately 1.5MB of raw data per second. The acquired signals undergo preprocessing, including DC component removal, anti-aliasing filtering, and digital low-pass filtering. The filter cutoff frequency is set to 8000Hz, and the filter order is 8.

[0079] In the signal processing stage, the technical team employed a fast independent component analysis (OCA) algorithm to separate multi-source coupled vibration signals. A 14×14 signal separation matrix was established to decompose the mixed vibration signal into 14 independent single-source vibration components. After 50 iterations, the signal separation convergence accuracy reached [percentage missing]. This effectively eliminated the mutual coupling interference between different vibration sources. The peak amplitude of the separated signal was reduced by approximately 28%, and the signal purity was significantly improved.

[0080] Wavelet transform time-frequency analysis was performed on the separated vibration signal, using the Daubechies-8 wavelet basis function for 5-level decomposition. Eight parameters were extracted from the vibration signal's time domain: mean, variance, peak value, RMS value, waveform factor, peak factor, impulse factor, and margin factor. Six parameters were extracted from the frequency domain: power spectral density, spectral centroid, band power ratio, and spectral entropy. Two parameters were extracted from the phase characteristics: phase difference and phase spectrum variance. Three parameters were extracted from the statistical characteristics: skewness, kurtosis, and fourth-order cumulant. A vibration feature vector set containing 19 feature parameters was established. Combining temperature signals with the analysis of vibration characteristic changes during the phase transition process, it was found that within ±15℃ of the phase transition temperature, the spectral centroid of the vibration signal shifts significantly by 10-15Hz.

[0081] The technical team analyzed the impact of changes in the physical state of phase change materials on their vibration characteristics using the phase change vibration mechanism equation. A reference density was set. 2100 Reference temperature At 20℃, the reference elastic modulus 25 Density variation coefficient The value is 0.85. Calculations show that at a phase transition temperature of 308℃, the material density change rate is -0.12, the elastic modulus change rate is -0.08, and the damping coefficient change rate is 0.23. A set of phase transition vibration characteristic parameters, including these three core parameters, was established.

[0082] The phase transition vibration feature recognition model employs a multi-layer feedforward neural network structure, comprising an input layer, three hidden layers, and an output layer. The input layer receives a 22-dimensional vector, including 19 vibration feature parameters and three phase transition vibration characteristic parameters. The first hidden layer contains 128 neurons and uses a sparse connection learning framework based on dynamic topology reconstruction. The initial connection density is 0.6, which stabilizes at 0.35 after training, reducing computational overhead by approximately 42%. The second and third hidden layers contain 64 and 32 neurons, respectively, and employ a fully connected structure. The output layer contains 8 neurons, corresponding to 8 different types of vibration anomalies, and uses a structured prediction framework based on a probabilistic graphical model. A belief propagation algorithm ensures the logical consistency of the prediction results.

[0083] During the establishment of the training dataset, the technical team collected six consecutive months of operational data from the thermal energy storage system, including 4200 hours of data under normal operating conditions and 380 hours of data under seven abnormal conditions. Abnormal conditions included bearing wear, imbalance, misalignment, loosening, resonance, pipe vibration, and abnormal phase change vibration. As shown in Table 1, training samples covering various operating conditions were established.

[0084] Table 1 Distribution of Vibration Anomaly Types in Thermal Energy Storage Systems

[0085] The dataset was divided into a training set of 5740 samples, a validation set of 1640 samples, and a test set of 820 samples in a 7:2:1 ratio. The model was trained using the Adam optimization algorithm with an initial learning rate of 0.001, a batch size of 64, and a weighted combination of cross-entropy loss and structured loss with a weight ratio of 0.7:0.3. After 500 iterations of training, the model achieved a recognition accuracy of 94.2% on the validation set and 92.8% on the test set.

[0086] The sparse connection learning framework based on dynamic topology reconstruction adaptively adjusts network parameters according to factors such as vibration signal complexity, sensor ratio, and phase transition state weights. The vibration signal complexity function calculates the current signal complexity index to be 0.73. The sensor ratio function, based on 12 effective sensors out of 14 total sensors, outputs a sparsity adjustment coefficient of 1.35. The phase transition state weight function, based on the current temperature of 320℃ and the phase transition temperature of 308℃, calculates a parameter weight coefficient of 1.15. The connection pruning threshold is set to 0.12, the sparsity target is set to 0.35, and the regularization weight is set to 0.005.

[0087] The structured prediction framework based on a probabilistic graphical model models the correlations among eight types of anomalous vibrations as an undirected graph. It performs five rounds of iterative calculations using the belief propagation algorithm and employs variational inference techniques to simplify the computational complexity of the joint probability distribution. The inference accuracy under the mean-field approximation assumption reaches [percentage missing]. This ensures the logical consistency of the prediction results and avoids the situation of predicting multiple contradictory anomalies at the same time.

[0088] The technical team established a vibration anomaly discrimination threshold system that includes multi-level discrimination criteria. For example... Figure 2 As shown, the vibration amplitude threshold is set according to different frequency bands; the threshold for the low-frequency band (0.1-10Hz) is 0.5. The threshold for the mid-frequency band (10-1000Hz) is 2.0. The threshold for the high-frequency band (1000-8000Hz) is 5.0. The frequency offset threshold is set to ±3Hz, and the phase change vibration abnormality probability threshold is set to 0.8. When the vibration characteristic parameters exceed the normal operating threshold range, or the phase change vibration abnormality probability exceeds 0.8, the system automatically triggers the abnormal vibration early warning mechanism.

[0089] As shown in Table 2, the key monitoring parameters are set during system operation: Table 2 System Monitoring Parameter Configuration Table

[0090] In actual operation, the system successfully detected multiple abnormal vibration events. One typical case occurred during the fourth month of operation, when abnormal vibration was caused by bearing wear in the molten salt pump. Figure 3 As shown, the system detected subtle changes in the vibration signal 72 hours before the anomaly occurred. The probability of abnormal phase change vibration gradually increased from the normal 0.1 to 0.85, triggering the early warning mechanism. Technicians performed timely maintenance, preventing further deterioration of the equipment failure.

[0091] The system also successfully identified abnormal vibration modes during the phase transition process. During the transition of molten salt from solid to liquid, abnormal phase transition vibrations caused by uneven local heating were detected. By analyzing the temperature vibration coupling signal and changes in phase transition vibration characteristic parameters, the system accurately identified the type of abnormal phase transition vibration with an anomaly probability score of 0.92, discovering potential problems 48 hours earlier than traditional methods.

[0092] Based on the type and severity of abnormal vibrations, the system adaptively adjusts the operating mode of the sensor array. When high-frequency mechanical vibration anomalies are detected, the system automatically increases the sampling frequency of the high-frequency sensors to 51200Hz, enhances the frequency selectivity of the filtering parameters, adjusts the pruning threshold of sparse connections in the neural network to 0.08, and improves the model's sensitivity to high-frequency anomaly features. When phase transition vibration anomalies are detected, the system increases the data weights of the temperature-vibration coupling sensors, adjusts the parameters of the phase transition state weight function, and optimizes the calculation accuracy of the phase transition vibration characteristic mechanism equation.

[0093] The entire monitoring system achieves 24-hour continuous online monitoring with a data processing latency of less than 200 milliseconds, an anomaly identification accuracy rate of over 92%, and a false alarm rate controlled within 3%. The system can simultaneously process data from up to 16 sensor channels, supporting the identification and classification of 8 different types of vibration anomalies, providing reliable technical support for the safe and stable operation of thermal energy storage equipment.

[0094] This invention represents a significant technological advancement over traditional vibration monitoring methods. Traditional methods typically monitor only a single vibration source, while this invention achieves effective separation of multi-source vibration signals through a multi-source vibration sensor array and independent component analysis (ICA) algorithms, resolving the signal coupling interference problem. Traditional methods struggle to handle complex vibration characteristic changes during phase transitions, while this invention establishes a quantitative relationship between physical parameter changes and vibration characteristics through the phase transition vibration characteristic mechanism equation, enabling accurate identification of abnormal phase transition vibrations. Traditional neural network models have fixed structures and high computational costs, while the dynamic topology reconstruction sparse connection learning framework employed in this invention adaptively adjusts the network structure according to signal complexity, significantly reducing computational complexity while maintaining prediction accuracy. Traditional methods may produce logical inconsistencies in their prediction results, while the probabilistic graphical model structured prediction framework used in this invention ensures logical consistency in multi-label prediction results, improving the reliability of anomaly type identification. These technological innovations enable this invention to achieve more accurate and efficient online monitoring and fault early warning of abnormal vibrations in thermal energy storage equipment.

[0095] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0096] Table 3. Variable Explanation Table (Part 1)

[0097] Table 4. Variable Explanation Table (Part Two)

[0098] 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. An online monitoring method for abnormal vibration of a thermal energy storage device, characterized in that, Multi-source vibration sensor arrays are deployed at key locations in thermal energy storage equipment to collect multi-channel vibration and temperature signals. These signals are then preprocessed and digitally filtered. A signal separation matrix is ​​established by separating multi-source coupled vibration signals based on independent component analysis (ICA). Wavelet transform time-frequency analysis is performed on the separated vibration signals to extract vibration feature vector sets. The influence of phase change material physical state changes on vibration characteristics is analyzed using the phase change vibration characteristic mechanism equation to establish a phase change vibration characteristic parameter set. An ICA characteristic identification model is used to process the vibration feature vector set, temperature data, and phase change vibration characteristic parameter set to identify abnormal vibration modes. A vibration anomaly discrimination threshold system is established to trigger an abnormal vibration early warning mechanism. The working mode of the sensor array and the signal processing algorithm parameters are adaptively adjusted according to the type and severity of the abnormal vibration.

2. The online monitoring method for abnormal vibration of thermal energy storage equipment according to claim 1, characterized in that, The arrangement steps of the multi-source vibration sensor array are as follows: on the thermal energy storage device, a combination of multiple vibration sensors of different types are arranged according to a preset spatial distribution pattern, including low-frequency vibration sensors and high-frequency vibration sensors. A differential arrangement method is used to reduce environmental noise interference. At the same time, a temperature vibration coupling sensor is set in the phase change material area to monitor the vibration status of the equipment in all directions.

3. The online monitoring method for abnormal vibration of thermal energy storage equipment according to claim 2, characterized in that, The differential arrangement method refers to installing sensors of the same type in pairs, and effectively suppressing common-mode noise interference through differential signal processing. The temperature-vibration coupling sensor refers to a composite sensor that simultaneously measures temperature and vibration, and is used to monitor the influence of temperature changes on vibration characteristics during phase transition.

4. The online monitoring method for abnormal vibration of thermal energy storage equipment according to claim 3, characterized in that, The acquisition steps for the multi-channel vibration and temperature signals specifically involve synchronous sampling using a high-precision data acquisition system. The sampling frequency is set to 2.56 times the highest frequency of the vibration signal according to the Nyquist sampling theorem. The acquired signals are then preprocessed and digitally filtered.

5. The online monitoring method for abnormal vibration of thermal energy storage equipment according to claim 4, characterized in that, The Independent Component Analysis (ICA) algorithm refers to a blind source separation algorithm that decomposes multiple mixed signals into statistically independent source signals based on the principle of statistical independence. It seeks a separation matrix that maximizes the statistical independence of the separated signals. The signal separation matrix refers to the linear transformation matrix used in the ICA algorithm to transform the mixed signals into independent signals.

6. The online monitoring method for abnormal vibration of thermal energy storage equipment according to claim 5, characterized in that, The wavelet transform time-frequency analysis refers to a signal analysis method that uses wavelet basis functions to decompose a signal into multiple scales and simultaneously obtain the signal's time and frequency information. The vibration feature vector set refers to a set of feature parameters representing the vibration state extracted from the vibration signal, including amplitude features, frequency features, phase features, and statistical features.

7. The online monitoring method for abnormal vibration of thermal energy storage equipment according to claim 6, characterized in that, The phase transition vibration characteristic mechanism equation is used to describe the influence of changes in physical parameters of the phase transition material on its vibration characteristics during the phase transition process. The inputs include the current temperature T and the phase transition temperature. The system calculates the material density ρ and elastic modulus E, outputs a set of phase transition vibration characteristic parameters, and calculates the rate of change of material density, rate of change of elastic modulus, and rate of change of damping coefficient during the phase transition process.

8. The online monitoring method for abnormal vibration of thermal energy storage equipment according to claim 7, characterized in that, The phase transition vibration characteristic parameter set refers to the set of parameters characterizing the changes in the physical properties of materials during the phase transition process, calculated by the phase transition vibration characteristic mechanism equation. These parameters include the material density change rate, the elastic modulus change rate, and the damping coefficient change rate. The material density change rate refers to the proportion of the material density change relative to the initial density during the phase transition process.

9. The online monitoring method for abnormal vibration of thermal energy storage equipment according to claim 8, characterized in that, The phase change vibration feature recognition model refers to a machine learning model used to identify the vibration feature change law during the phase change process of thermal energy storage materials. It learns the mapping relationship between the phase change process and vibration features through training, and outputs the phase change vibration anomaly probability and anomaly type identifier. The value range of the phase change vibration anomaly probability is ∈ [0, 1].

10. The online monitoring method for abnormal vibration of thermal energy storage equipment according to claim 9, characterized in that, The phase change vibration feature recognition model is structured as a multi-layer feedforward neural network, which includes an input layer, three hidden layers, and an output layer. The input layer receives a set of vibration feature vectors, temperature data, and a set of phase change vibration characteristic parameters. The first hidden layer adopts a sparse connection learning framework based on dynamic topology reconstruction, the second and third hidden layers adopt a fully connected structure, and the output layer adopts a structured prediction framework based on a probabilistic graphical model.

Images