Fault judgment method of fluidized bed energy storage system

Abstract

The invention provides a fault judgment method for a fluidized bed energy storage system, and belongs to the technical field of fluidized bed energy storage systems. Periodic pressure pulses are generated by using a vibration excitation source at the bottom of a bed layer, and propagation characteristics of the periodic pressure pulses in the bed layer are detected to form a pressure fluctuation matrix; the method comprises the following steps of: performing multi-scale decomposition on parameters such as bed pressure drop gradient, particle concentration, heat transfer coefficient and bed void ratio by adopting a wavelet transform algorithm to establish a frequency domain gain matrix; analyzing a bed fluidization state through a flow state identification model, and constructing a fault symptom identifier based on a support vector machine algorithm to extract frequency domain energy distribution characteristics and wave impedance statistical characteristics to calculate a fault probability value; and when the fault probability value exceeds a threshold value, establishing an abnormal state matrix, and performing comprehensive evaluation by using an abnormal state evaluation function to output a final fault judgment result. The technical problem that the internal structure of the fluidized bed energy storage system cannot be accurately identified and positioned in an early stage is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of fluidized bed energy storage systems, and more specifically, relates to a fault determination method for fluidized bed energy storage systems. Background Technology

[0002] Fluidized bed energy storage systems, as an emerging energy storage technology, are widely used in industrial waste heat recovery, renewable energy storage, and grid peak shaving. Their core lies in achieving efficient energy storage and release through the flow and heat transfer characteristics of the fluidized medium. Traditional fault detection methods mainly rely on single-parameter monitoring methods such as temperature and pressure sensors, using threshold alarms for fault diagnosis. This method is widely used in practical applications such as thermal power plant thermal storage systems, fluidized bed reactors in the chemical industry, and biomass fluidized bed boilers. However, current fault detection technologies for fluidized bed energy storage systems only acquire parameter information from the bed surface or local areas, failing to deeply analyze the three-dimensional spatial structural changes within the bed. This results in a lack of effective early identification capabilities for internal structural anomalies such as particle aggregation, void formation, and uneven fluidization. In other words, existing technologies suffer from the technical problem of failing to achieve early and accurate identification and location of internal structural anomalies in fluidized bed energy storage systems. Summary of the Invention

[0003] In view of this, the present invention provides a fault determination method for fluidized bed energy storage systems, which can solve the technical problem in the prior art that it is impossible to achieve early and accurate identification and location of internal structural anomalies in fluidized bed energy storage systems.

[0004] This invention is implemented as follows: A fault determination method for a fluidized bed energy storage system is provided. Pressure sensor arrays are arranged at different heights within the fluidized bed, while temperature and flow sensors are installed on the bed sidewalls to collect basic monitoring parameters and establish an operating parameter matrix. Periodic pressure pulses are generated by a vibration excitation source installed at the bottom of the bed. The pressure sensor array detects the propagation speed and attenuation characteristics of the pressure pulses in different bed regions, calculates the wave impedance value at each monitoring point, and marks an abnormal point when the wave impedance change rate exceeds a set threshold, establishing a pressure fluctuation matrix. State feature parameters are extracted, and multi-scale decomposition of these parameters is performed using a wavelet transform algorithm to obtain fault symptom signals in different frequency bands, establishing a frequency domain gain matrix. The operating parameter matrix is ​​then input into a flow regime identification model. The fluidization state of the bed is analyzed, and the fluidization state assessment value is output by the fluidization identification model. A fault symptom identifier is established based on the support vector machine algorithm. The main frequency band energy distribution features are extracted from the frequency domain gain matrix, and the wave impedance statistical features are extracted from the pressure fluctuation matrix. The main frequency band energy distribution features and wave impedance statistical features are combined to form an input feature vector. The fault probability value is calculated by the fault symptom identifier. When the fault probability value exceeds the threshold, an abnormal state matrix is ​​established. The abnormal state matrix is ​​comprehensively evaluated using an abnormal state evaluation function. The abnormal state evaluation function outputs the final fault judgment result. Corresponding handling measures are formulated based on the final fault judgment result. A fault early warning mechanism is established. When a potential fault symptom is detected, the early warning program is activated, and the final fault judgment result and handling suggestions are output to the monitoring system interface.

[0005] Specifically, the establishment of the operating condition parameter matrix involves a two-dimensional data matrix formed by arranging basic monitoring parameters such as bed temperature, fluidizing velocity, pressure drop, gas flow rate, and bed height in a time sequence. The bed temperature is obtained by thermocouple temperature sensors installed at different heights of the bed, the fluidizing velocity is measured by a turbine flow meter installed in the inlet duct, the pressure drop is measured by differential pressure sensors installed at the bottom and top of the bed, the gas flow rate is measured by an orifice plate flow meter in the inlet duct, and the bed height is measured by an ultrasonic level gauge installed on the side wall of the bed.

[0006] Specifically, the pressure fluctuation matrix is ​​established by obtaining a pressure pulse propagation characteristic data matrix inside the bed through a pressure pulse propagation detection algorithm. This matrix includes wave impedance values, wave velocity change rate, and attenuation coefficient parameters for each monitoring point. The wave impedance value is calculated by multiplying the pressure pulse propagation velocity and the medium density. The wave velocity change rate is calculated by the ratio of the difference between the current wave velocity and the reference wave velocity to the reference wave velocity. The attenuation coefficient is obtained by exponential fitting of the pressure pulse amplitude attenuation law.

[0007] Specifically, the frequency domain gain matrix is ​​established by performing multi-scale decomposition of the bed pressure drop gradient, particle concentration, heat transfer coefficient, and bed porosity state characteristic parameters using a wavelet transform algorithm. The bed pressure drop gradient is calculated by the ratio of the pressure difference to the distance between adjacent pressure measurement points. The particle concentration is obtained by online measurement using an optical concentration meter installed in the observation window on the bed sidewall. The heat transfer coefficient is calculated by the ratio of the heat flux density on the bed wall to the temperature difference. The bed porosity is calculated by the ratio of the difference between the bed volume and the total volume of solid particles to the bed volume.

[0008] The fluidization identification model refers to a bed fluidization state identification model based on a liquid neural network architecture, which includes three main parts: an input layer, a liquid computation layer, and a readout layer. The liquid computation layer adopts a sparsely connected dynamic neural network structure, and the connection weights between neurons are dynamically adjusted according to the bed temperature and fluidization velocity. The model captures the dynamic characteristics of the bed fluidization process through a temporal memory mechanism, and the attention weight parameters of the sparse attention mechanism are adaptively adjusted according to the pressure drop rate.

[0009] Specifically, the step of establishing the training dataset for the fluidized bed recognition model involves collecting fluidized bed operation data under different working conditions, including three typical working conditions: normal fluidization state, non-uniform fluidization state, and abnormal fluidization state. More than 1,000 sets of multi-dimensional time-series data are collected for each working condition. The data acquisition frequency is 10Hz, and the single acquisition duration is 300s. Noise and outliers are eliminated through data preprocessing. The continuous time-series data is divided into fixed-length training samples using the sliding window method. Each training sample contains 30s of time-series data.

[0010] The training steps of the fluid recognition model specifically involve dividing the dataset into 70% for training, 20% for validation, and 10% for testing. An adaptive learning rate optimization algorithm is used to iteratively update the model parameters. During training, the model performance is evaluated using cross-validation. The convergence condition is set to a change in the validation set loss function of less than 0.001 within 20 consecutive training cycles. After training, the generalization ability of the model is evaluated using the test set to ensure that the recognition accuracy of the model under different working conditions exceeds 95%.

[0011] Among them, a fluidization state assessment value in the range of 0 to 0.35 is considered a normal fluidization state, a fluidization state assessment value in the range of 0.35 to 0.75 is considered a non-uniform fluidization state requiring further monitoring, and a fluidization state assessment value in the range of 0.75 to 1 is considered an abnormal fluidization state.

[0012] Specifically, the training dataset establishment step of the fault symptom identifier involves collecting fluidized bed operation data under different fault types and normal conditions, including five typical states: blockage fault state, void fault state, fluidization non-uniformity fault state, heat transfer abnormality fault state, and normal operation state. More than 500 sets of multi-dimensional monitoring data are collected for each state, with a data acquisition time of 600s per set and an acquisition frequency of 5Hz. The main frequency band energy distribution characteristics and wave impedance statistical characteristics are extracted from each set of data to form a feature vector.

[0013] The main frequency band energy distribution features include three components: the energy proportion of the low frequency band (0.1Hz to 1Hz), the energy proportion of the mid frequency band (1Hz to 10Hz), and the energy proportion of the high frequency band (10Hz to 50Hz). The wave impedance statistical features include four components: the wave impedance mean, the wave impedance standard deviation, the maximum wave impedance change rate, and the minimum wave impedance change rate. The seven-dimensional feature vector is combined with the corresponding fault type label to establish a supervised learning training dataset. Data standardization is used to eliminate differences in the dimensions of different features.

[0014] The basic monitoring parameters include bed temperature, fluidizing velocity, pressure drop, gas flow rate, and bed height. State characteristic parameters include bed pressure drop gradient, particle concentration, heat transfer coefficient, and bed porosity.

[0015] Monitoring points refer to the installation locations of various sensors arranged at specific positions and intervals within the fluidized bed energy storage system. These primarily include differential pressure sensor arrays spaced vertically every 0.2 meters inside the bed, thermocouple temperature sensor monitoring points installed on the bed sidewalls at 25%, 50%, and 75% of the bed height, and flow sensor monitoring points installed in the air inlet duct. These monitoring points form a comprehensive monitoring network using multi-sensor fusion technology, collecting real-time data on fundamental parameters such as bed temperature, fluidizing velocity, pressure drop, gas flow rate, and bed height. Simultaneously, pressure sensor arrays detect the propagation characteristics of pressure pulses, calculating parameters such as wave impedance, wave velocity change rate, and attenuation coefficient for each monitoring point. This provides comprehensive data support for fault diagnosis and condition assessment of the fluidized bed energy storage system.

[0016] Specifically, the training steps of the fault symptom recognizer involve dividing the training dataset into a training set and a test set in an 8:2 ratio, using a radial basis function as the kernel function of the support vector machine, optimizing the kernel function parameters and penalty parameters using a grid search method, with the kernel function parameters ranging from 0.001 to 10 and the penalty parameters ranging from 1 to 1000, evaluating the classification performance of different parameter combinations using a 10-fold cross-validation method, selecting the parameter combination with the highest cross-validation accuracy as the optimal parameters, and using the optimal parameters to train the final fault symptom recognizer model.

[0017] The abnormal state matrix refers to a data matrix formed by organizing the corresponding abnormal monitoring parameter information according to spatial location and time sequence when the fault probability value output by the fault symptom identifier exceeds a set threshold. The spatial location information of the abnormal state matrix is ​​determined by the installation location coordinates of the pressure sensor array, and the time information is determined by the timestamp record of the data acquisition system.

[0018] Specifically, the abnormal state evaluation function performs a comprehensive analysis of the abnormal state matrix and assesses the severity of the fault. The inputs include the abnormal state matrix, pressure fluctuation matrix, frequency domain gain matrix, historical fault data, and current operating parameters. The outputs are the fault type classification results and the fault severity level. The fault type classification results include four types: blockage fault, void fault, fluidization non-uniformity fault, and heat transfer abnormal fault. The fault severity level is divided into three levels: minor fault, moderate fault, and severe fault.

[0019] Specifically, the pressure pulse propagation detection algorithm is based on the principle of acoustic propagation. It generates periodic pressure pulses by installing a vibration excitation source at the bottom of the fluidized bed. The propagation process of the pressure pulses is monitored by an array of pressure sensors arranged at different positions in the bed. The vibration excitation source generates periodic vibrations with a frequency of 1Hz to 100Hz and a vibration amplitude of 0.1mm to 5mm through an electromagnetic drive mechanism. The algorithm establishes a threshold judgment mechanism to identify abnormal areas by calculating the wave impedance change rate of each monitoring point in real time.

[0020] Among them, the energy proportion in the low-frequency band from 0.1Hz to 1Hz reflects the large-scale fluidization fluctuation characteristics of the bed, the energy proportion in the mid-frequency band from 1Hz to 10Hz reflects the medium-scale particle motion characteristics of the bed, the energy proportion in the high-frequency band from 10Hz to 50Hz reflects the small-scale turbulence characteristics of the bed, the mean wave impedance reflects the overall density level of the bed, the standard deviation of wave impedance reflects the uniformity of the density distribution of the bed, and the maximum and minimum values ​​of the wave impedance change rate reflect the degree of abrupt change in the bed structure.

[0021] The threshold value for the fault probability is 0.6. Under normal fluidization conditions, the propagation speed of the pressure pulse is relatively stable and the attenuation law is in line with theoretical expectations. In abnormal areas, the change in medium density and elastic modulus will cause a sudden change in wave impedance. By analyzing the wave velocity distribution spectrum and attenuation coefficient, the abnormality of the internal structure of the bed can be accurately located, providing technical support for the preventive maintenance of fluidized bed energy storage systems.

[0022] This invention constructs a multi-dimensional fault diagnosis system based on pressure pulse propagation detection, establishing a comprehensive analysis framework including a pressure fluctuation matrix, a frequency domain gain matrix, and an abnormal state matrix, thereby achieving precise monitoring and identification of three-dimensional spatial structural changes within the fluidized bed. This invention utilizes the differences in the propagation characteristics of periodic pressure pulses generated by vibration excitation sources within the bed, combined with real-time detection of wave impedance changes by a pressure sensor array, effectively overcoming the technical limitations of traditional single-parameter monitoring methods that cannot penetrate deep into the bed. Through the fusion application of multi-scale wavelet transform and support vector machine algorithms, the early identification accuracy and location accuracy of fault symptoms are significantly improved. In summary, this invention solves the technical problem mentioned in the background art of the inability to achieve early and accurate identification and location of internal structural anomalies in fluidized bed energy storage systems. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention.

[0024] Figure 2 This is a schematic diagram of the network structure of the flow recognition model involved in the present invention. Detailed Implementation

[0025] 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.

[0026] like Figure 1 The diagram shown is a flowchart of a fault determination method for a fluidized bed energy storage system provided by the present invention. This method includes the following steps: S01. Arrange pressure sensor arrays at different heights inside the fluidized bed energy storage system, and install temperature and flow sensors on the sidewalls of the bed to collect bed temperature, fluidization velocity, pressure drop, gas flow rate, and bed height as basic monitoring parameters to establish an operating parameter matrix. S02. Periodic pressure pulses are generated by a vibration excitation source installed at the bottom of the bed. The propagation speed and attenuation characteristics of the pressure pulses in different bed areas are detected by a pressure sensor array. The wave impedance value of each monitoring point is calculated. When the wave impedance change rate exceeds the set threshold, it is marked as an abnormal point. A pressure fluctuation matrix is ​​established. S03. Extract the bed pressure drop gradient, particle concentration, heat transfer coefficient, and bed porosity as state characteristic parameters. Perform multi-scale decomposition on the state characteristic parameters using wavelet transform algorithm to obtain fault symptom signals in different frequency bands and establish a frequency domain gain matrix. S04. Input the working condition parameter matrix into the fluidization identification model to analyze the fluidization state of the bed. The fluidization identification model outputs the fluidization state evaluation value of the bed. When the fluidization state evaluation value of the bed is in the range of [0, 0.35), it is determined to be a normal fluidization state. When the fluidization state evaluation value of the bed is in the range of [0.35, 0.75), it is determined to be a non-uniform fluidization state that needs further monitoring. When the fluidization state evaluation value of the bed is in the range of [0.75, 1], it is determined to be an abnormal fluidization state. S05. A fault symptom identifier is established based on the support vector machine algorithm. The main frequency band energy distribution features are extracted from the frequency domain gain matrix, and the wave impedance statistical features are extracted from the pressure fluctuation matrix. The main frequency band energy distribution features and wave impedance statistical features are combined to form an input feature vector. The fault probability value is calculated through the fault symptom identifier. When the fault probability value exceeds 0.6, an abnormal state matrix is ​​established. S06. Use the abnormal state evaluation function to comprehensively evaluate the abnormal state matrix. The abnormal state evaluation function outputs the final fault judgment result. Based on the final fault judgment result, formulate corresponding handling measures. S07. Establish a fault early warning mechanism. When potential fault signs are detected, the early warning procedure is activated, and the final fault judgment result and handling suggestions are output to the monitoring system interface.

[0027] The operating parameter matrix is ​​a two-dimensional data matrix formed by arranging basic monitoring parameters such as bed temperature, fluidizing velocity, pressure drop, gas flow rate, and bed height in a time series. It is used to describe the real-time operating status of the fluidized bed energy storage system. The bed temperature is obtained through thermocouple temperature sensors installed at different heights within the bed, with a measurement accuracy of ±1℃ and a measurement range of 20℃ to 800℃. The fluidizing velocity is measured by a turbine flow meter installed in the inlet duct, with a measurement range of 0.1 m / s to 10 m / s and a measurement accuracy of ±0.05 m / s. The pressure drop is measured by differential pressure sensors installed at the bottom and top of the bed, with a measurement accuracy of ±0.1 kPa and a measurement range of 0 kPa to 50 kPa. The gas flow rate is measured by an orifice plate flow meter in the inlet duct, with a measurement accuracy of ±2% and a measurement range of 0... / h to 1000 / h. The bed height is measured by an ultrasonic level gauge installed on the side wall of the bed, with a measurement accuracy of ±5mm and a measurement range of 0.5m to 3m.

[0028] The frequency domain gain matrix is ​​a frequency domain feature matrix obtained by multi-scale decomposition of state characteristic parameters such as bed pressure drop gradient, particle concentration, heat transfer coefficient, and bed porosity using a wavelet transform algorithm. It is used to identify fault symptom signals in different frequency bands. The bed pressure drop gradient is calculated as the ratio of the pressure difference to the distance between adjacent pressure measuring points. The pressure difference between adjacent pressure measuring points is measured by a pressure sensor array, and the distance between adjacent pressure measuring points is a fixed installation spacing of 0.2m. The particle concentration is measured online using an optical concentration meter installed in the observation window on the bed sidewall. The measurement principle is based on the relationship between light scattering intensity and particle concentration, with a measurement accuracy of ±5% and a measurement range of 0 kg / m³. Up to 800kg / The heat transfer coefficient is calculated as the ratio of heat flux density at the bed wall to the temperature difference. The heat flux density is measured by a heat flux meter installed on the bed wall with an accuracy of ±3%, and the temperature difference is measured by a thermocouple temperature sensor. The bed porosity is calculated as the ratio of the difference between the bed volume and the total volume of solid particles to the bed volume. The bed volume is calculated from the bed height and bed cross-sectional area, and the total volume of solid particles is calculated from the particle mass and particle density. The particle mass is measured by a weighing sensor installed in the feeding system.

[0029] The abnormal state matrix is ​​a data matrix formed by organizing the corresponding abnormal monitoring parameter information according to spatial location and time sequence when the fault probability value output by the fault symptom identifier exceeds a set threshold. It is used to locate the location and time of the fault. The spatial location information of the abnormal state matrix is ​​determined by the installation coordinates of the pressure sensor array, and the time information is determined by the timestamp records of the data acquisition system.

[0030] The pressure fluctuation matrix is ​​a data matrix of pressure pulse propagation characteristics within the bed, obtained through a pressure pulse propagation detection algorithm. It includes parameters such as wave impedance values, wave velocity change rate, and attenuation coefficient at each monitoring point, and is used to identify structural anomalies within the bed. The wave impedance value is calculated by multiplying the pressure pulse propagation velocity by the medium density. The pressure pulse propagation velocity is calculated from the signal propagation time difference of the pressure sensor array, and the medium density is calculated by a weighted average of particle concentration and gas density. The wave velocity change rate is calculated as the ratio of the difference between the current wave velocity and the reference wave velocity to the reference wave velocity. The attenuation coefficient is obtained by exponential fitting of the pressure pulse amplitude attenuation law.

[0031] The fluidization identification model is a fluidized bed state identification model based on a liquid neural network architecture, comprising three main parts: an input layer, a liquid computation layer, and a readout layer. The liquid computation layer employs a sparsely connected dynamic neural network structure, with the connection weights between neurons dynamically adjusted according to bed temperature and fluidization velocity. The model captures the dynamic characteristics of the bed fluidization process through a temporal memory mechanism, and the attention weight parameters of the sparse attention mechanism are adaptively adjusted based on the pressure drop rate. The steps for establishing the training dataset for the fluidization identification model include collecting fluidized bed operation data under different operating conditions, including three typical operating conditions: normal fluidization, non-uniform fluidization, and abnormal fluidization. More than 1000 sets of multi-dimensional time-series data are collected for each operating condition, with a data acquisition frequency of 10Hz and a single acquisition duration of 300s. Noise and outliers are eliminated through data preprocessing. A sliding window method is used to divide the continuous time-series data into fixed-length training samples, each containing 30s of time-series data, thus establishing a supervised learning dataset containing input feature vectors and corresponding labels. The training steps of the fluid recognition model include dividing the dataset into 70% for training, 20% for validation, and 10% for testing; using an adaptive learning rate optimization algorithm to iteratively update the model parameters; evaluating model performance during training using cross-validation; setting the convergence condition as a change in the validation set loss function of less than 0.001 within 20 consecutive training cycles; and evaluating the model's generalization ability using the test set after training to ensure that the model's recognition accuracy exceeds 95% under different working conditions.

[0032] The abnormal state evaluation function is used to comprehensively analyze the abnormal state matrix and assess the severity of faults. Inputs include the abnormal state matrix, pressure fluctuation matrix, frequency domain gain matrix, historical fault data, and current operating parameters. Outputs are the fault type classification results and the fault severity level. The historical fault data is obtained from historical fault records stored in the fault database, including fault occurrence time, fault type, fault location, and handling measures. The current operating parameters are provided by the operating parameter matrix. The fault type classification results include four types: blockage faults, void faults, fluidization non-uniformity faults, and heat transfer abnormality faults. The fault severity level is divided into three levels: minor faults, moderate faults, and severe faults.

[0033] The energy distribution characteristics of the main frequency bands were obtained by dividing the frequency domain gain matrix into frequency bands and performing energy statistical calculations. The energy proportion in the low-frequency band (0.1Hz to 1Hz) reflects the large-scale fluidization fluctuation characteristics of the bed, the energy proportion in the mid-frequency band (1Hz to 10Hz) reflects the medium-scale particle motion characteristics of the bed, and the energy proportion in the high-frequency band (10Hz to 50Hz) reflects the small-scale turbulence characteristics of the bed. The wave impedance statistical characteristics were obtained through statistical analysis of the pressure fluctuation matrix. The mean wave impedance reflects the overall density level of the bed, the standard deviation of the wave impedance reflects the uniformity of the bed density distribution, and the maximum and minimum values ​​of the wave impedance change rate reflect the degree of abrupt changes in the bed structure.

[0034] The steps for establishing the training dataset for the fault symptom identifier include collecting fluidized bed operation data under different fault types and normal conditions, including five typical states: blockage fault state, void fault state, fluidization non-uniformity fault state, heat transfer abnormality fault state, and normal operation state. More than 500 sets of multidimensional monitoring data are collected for each state, with a data acquisition time of 600s per set and an acquisition frequency of 5Hz. From each set of data, the main frequency band energy distribution features and wave impedance statistical features are extracted to form a feature vector. The main frequency band energy distribution features include three components: the energy proportion of the low frequency band (0.1Hz to 1Hz), the energy proportion of the mid frequency band (1Hz to 10Hz), and the energy proportion of the high frequency band (10Hz to 50Hz). The wave impedance statistical features include four components: wave impedance mean, wave impedance standard deviation, maximum wave impedance change rate, and minimum wave impedance change rate. The seven-dimensional feature vector is combined with the corresponding fault type label to establish a supervised learning training dataset, and the differences in different feature dimensions are eliminated through data standardization processing. The training steps of the fault symptom recognizer include dividing the training dataset into a training set and a test set in an 8:2 ratio, using a radial basis function as the kernel function of the support vector machine, optimizing the kernel function parameters and penalty parameters using a grid search method, with the kernel function parameters ranging from 0.001 to 10 and the penalty parameters ranging from 1 to 1000, evaluating the classification performance of different parameter combinations using a 10-fold cross-validation method, selecting the parameter combination with the highest cross-validation accuracy as the optimal parameters, training the final fault symptom recognizer model using the optimal parameters, and validating the model's generalization performance using a test set to ensure that the fault symptom recognizer achieves a recognition accuracy of over 90% under various fault conditions.

[0035] The pressure pulse propagation detection algorithm is based on the principle of acoustic propagation. It generates periodic pressure pulses by installing a vibration excitation source at the bottom of the fluidized bed, and monitors the propagation process of these pulses using an array of pressure sensors placed at different locations within the bed. The vibration excitation source generates periodic vibrations with frequencies ranging from 1Hz to 100Hz and amplitudes from 0.1mm to 5mm via an electromagnetic drive mechanism. The core of the algorithm lies in analyzing the changes in the propagation velocity and attenuation characteristics of the pressure pulses within the bed. When abnormal conditions such as particle aggregation, void formation, or uneven fluidization occur within the bed, the density distribution of the bed will change locally, leading to significant differences in the propagation characteristics of the pressure pulses. Under normal fluidization conditions, the pressure pulse propagation velocity is relatively stable, and the attenuation pattern conforms to theoretical expectations. However, in abnormal areas, changes in medium density and elastic modulus cause abrupt changes in wave impedance. The algorithm calculates the wave impedance change rate at each monitoring point in real time, establishes a threshold judgment mechanism to identify abnormal areas, and combines wave velocity distribution maps and attenuation coefficient analysis to achieve precise location of structural anomalies within the bed. Compared with traditional temperature and pressure monitoring, this method has higher sensitivity and timeliness, and can capture minute structural change signals in the early stages of a fault. It provides important technical support for preventive maintenance of fluidized bed energy storage systems and significantly improves the safety and reliability of system operation.

[0036] The steps for establishing the training dataset for the fault symptom identifier include collecting fluidized bed operation data under different fault types and normal conditions, including five typical states: blockage fault state, void fault state, fluidization non-uniformity fault state, heat transfer abnormality fault state, and normal operation state. More than 500 sets of multidimensional monitoring data are collected for each state, with a data acquisition time of 600s per set and an acquisition frequency of 5Hz. From each set of data, the main frequency band energy distribution features and wave impedance statistical features are extracted to form a feature vector. The main frequency band energy distribution features include three components: the energy proportion of the low frequency band (0.1Hz to 1Hz), the energy proportion of the mid frequency band (1Hz to 10Hz), and the energy proportion of the high frequency band (10Hz to 50Hz). The wave impedance statistical features include four components: wave impedance mean, wave impedance standard deviation, maximum wave impedance change rate, and minimum wave impedance change rate. The seven-dimensional feature vector is combined with the corresponding fault type label to establish a supervised learning training dataset, and the differences in different feature dimensions are eliminated through data standardization processing. The training steps of the fault symptom recognizer include dividing the training dataset into a training set and a test set in an 8:2 ratio, using a radial basis function as the kernel function of the support vector machine, optimizing the kernel function parameters and penalty parameters using a grid search method, with the kernel function parameters ranging from 0.001 to 10 and the penalty parameters ranging from 1 to 1000, evaluating the classification performance of different parameter combinations using a 10-fold cross-validation method, selecting the parameter combination with the highest cross-validation accuracy as the optimal parameters, training the final fault symptom recognizer model using the optimal parameters, and validating the model's generalization performance using a test set to ensure that the fault symptom recognizer achieves a recognition accuracy of over 90% under various fault conditions.

[0037] The specific implementation methods of the above steps are described in detail below.

[0038] The specific implementation of step S01 involves establishing a comprehensive monitoring network for the fluidized bed energy storage system using multi-sensor fusion technology. This step aims to acquire basic system operation data and construct a time-series operating parameter matrix. First, differential pressure sensor arrays are arranged vertically at 0.2m intervals inside the bed. The sensors employ a ceramic diaphragm structure to adapt to high-temperature environments, achieving a measurement accuracy of ±0.1kPa and a measurement range covering 0kPa to 50kPa. Next, K-type thermocouple temperature sensors are installed at equal intervals on the bed sidewalls, with measurement points distributed at 25%, 50%, and 75% of the bed height. The sensor protective sleeves are made of stainless steel, achieving a measurement accuracy of ±1℃ and a measurement range of 20℃ to 800℃. Then, a turbine flow meter is installed in the inlet duct to measure the fluidizing velocity. The flow meter impeller is made of titanium alloy, with a measurement range of 0.1m / s to 10m / s and a measurement accuracy of ±0.05m / s. Simultaneously, an orifice plate flow meter is installed in the inlet duct to measure the gas flow rate. The orifice plate has an opening ratio of 0.6 and a measurement range of 0. / h to 1000 / h, measurement accuracy ±2%. Finally, an ultrasonic level gauge is installed on the side wall of the bed to measure the bed height. The ultrasonic frequency is 40kHz, the measurement range is 0.5m to 3m, and the measurement accuracy is ±5mm. The data acquisition system synchronously acquires all sensor signals at a frequency of 10Hz, and uses a moving average filtering algorithm to eliminate high-frequency noise. The filtering window length is set to 5 sampling points. The operating parameter matrix organizes the data according to two dimensions: time series and spatial location. The matrix rows represent the time series, and the columns represent different monitoring parameters. Each row contains five basic parameters: bed temperature, fluidizing velocity, pressure drop, gas flow rate, and bed height, which are used to describe the complete operating status of the fluidized bed energy storage system at a certain time.

[0039] The specific implementation of step S02 is based on establishing an internal structural anomaly detection mechanism for the bed layer according to the principle of acoustic wave propagation. This step identifies local structural changes by analyzing the propagation characteristics of pressure pulses within the bed layer. First, an electromagnetic vibrator is installed at the center of the bed bottom as the vibration excitation source. The vibrator drives a magnetic piston via an electromagnetic coil to generate periodic pressure pulses. The pulse frequency range is 1Hz to 100Hz, and the vibration amplitude is 0.1mm to 5mm. The excitation signal uses a frequency sweep method to cover different frequency bands. Then, the propagation process of the pressure pulses is detected using an array of pre-placed pressure sensors. The sensor sampling frequency is set to 1000Hz to ensure the complete waveform of the pulse signal is captured. Next, a cross-correlation algorithm is used to calculate the signal propagation time difference between adjacent sensors. The pressure pulse propagation velocity is obtained by the ratio of the propagation distance to the time difference. The product of the pressure pulse propagation velocity and the medium density is the wave impedance value. The medium density is calculated using a volume-weighted average method of particle concentration and gas density. The ratio of the difference between the current wave velocity and the reference wave velocity to the reference wave velocity is defined as the wave velocity change rate. The reference wave velocity is the average propagation velocity under normal fluidization conditions. The rate of change of wave impedance is calculated as the ratio of the difference between the current wave impedance and the historical average wave impedance to the historical average wave impedance. An outlier is identified when the rate of change exceeds 15%. The attenuation law of the pressure pulse amplitude is obtained by fitting an exponential attenuation model, with parameters optimized using the least squares method. The pressure fluctuation matrix organizes data according to both spatial location and frequency, including parameters such as wave impedance value, wave velocity change rate, and attenuation coefficient at each monitoring point, used to quantify changes in the acoustic propagation characteristics within the bed.

[0040] The specific implementation of step S03 is based on wavelet transform multi-scale signal decomposition technology to extract the frequency domain features of the bed state. This step converts the time-domain signal into frequency-domain information to identify abnormal signs at different scales. First, the bed pressure drop gradient is calculated by dividing the pressure difference between adjacent pressure measuring points by the distance between the measuring points. The distance between the measuring points is a fixed installation spacing of 0.2m. The pressure drop gradient reflects the pressure distribution characteristics of the bed along the height direction. Then, the particle concentration is measured online using a laser scattering optical concentration meter installed in the observation window on the side wall of the bed. The measurement principle is based on Mie scattering theory, with an incident laser wavelength of 632.8nm, a scattering angle of 90°, a measurement accuracy of ±5%, and a measurement range of 0 kg / m³. Up to 800kg / Next, the heat transfer coefficient was calculated using the ratio of bed wall heat flux density to temperature difference. Heat flux density was measured using a thin-film heat flux meter with an accuracy of ±3%, and temperature difference was obtained from adjacent thermocouple temperature sensors. Bed porosity was calculated by dividing the difference between bed volume and the total volume of solid particles by the bed volume. Bed volume was obtained by multiplying bed height by bed cross-sectional area, and the total volume of solid particles was obtained by dividing particle mass by particle density. Particle mass was measured by strain gauge weighing sensors in the feeding system. De Bessie wavelet was used as the mother wavelet function to perform multi-scale wavelet decomposition of bed pressure drop gradient, particle concentration, heat transfer coefficient, and bed porosity. The decomposition layer was set to 6 layers to obtain wavelet coefficients for different frequency bands. The energy distribution of the wavelet coefficients was calculated using Passevar's theorem to determine the energy proportion of each frequency band. The frequency domain gain matrix organized the data according to two dimensions: frequency band and signal type, to characterize the energy distribution characteristics of fault symptom signals in different frequency bands.

[0041] The specific implementation of step S04 involves constructing an intelligent identification system for bed fluidization state using a liquid neural network architecture. This step automatically identifies and quantitatively evaluates the fluidization state of the bed through a deep learning model. The fluidization identification model employs a liquid neural network structure, comprising three main parts: an input layer, a liquid computation layer, and a readout layer. The input layer receives the 5-dimensional feature vector of the operating condition parameter matrix. The liquid computation layer uses a sparsely connected dynamic neural network with 200 neurons and a connection sparsity of 20%. The connection weights between neurons are dynamically adjusted based on changes in bed temperature and fluidization velocity, using a gradient descent optimization method. The temporal memory mechanism of the liquid computation layer is implemented through cyclic connections, with a memory length of 30 time steps, used to capture the dynamic characteristics and temporal dependencies of the bed fluidization process. The attention weight parameters of the sparse attention mechanism are adaptively adjusted based on the pressure drop rate, and the attention weight calculation uses soft maximum function normalization. The readout layer employs a fully connected structure with 50 neurons and a modified linear unit activation function. The output layer uses a sigmoid function to limit the output value to between 0 and 1 as the evaluation value for bed fluidization status. A fluidization status evaluation value between 0 and 0.35 is considered normal fluidization; between 0.35 and 0.75 indicates non-uniform fluidization requiring further monitoring; and between 0.75 and 1 indicates abnormal fluidization requiring immediate intervention.

[0042] The specific implementation of step S05 is to establish a multi-feature fusion fault symptom identification system based on the support vector machine algorithm. The function of this step is to automatically identify and classify different types of fault symptoms through machine learning methods. First, the energy distribution features of the main frequency bands are extracted from the frequency domain gain matrix, and the spectrum is divided into three intervals: 0.1Hz to 1Hz (low frequency band), 1Hz to 10Hz (mid frequency band), and 10Hz to 50Hz (high frequency band). The energy proportion of each frequency band is calculated. The energy proportion of the low frequency band reflects the large-scale fluidization fluctuation characteristics of the bed, the energy proportion of the mid frequency band reflects the medium-scale particle motion characteristics of the bed, and the energy proportion of the high frequency band reflects the small-scale turbulence characteristics of the bed. Then, the wave impedance statistical features are extracted from the pressure fluctuation matrix, including four statistical quantities: mean wave impedance, standard deviation of wave impedance, maximum wave impedance change rate, and minimum wave impedance change rate. The mean wave impedance reflects the overall density level of the bed, the standard deviation of wave impedance reflects the uniformity of the bed density distribution, and the extreme value of the wave impedance change rate reflects the degree of abrupt change in the bed structure. The energy proportion features of three frequency bands and the statistical features of wave impedance are combined to form a 7-dimensional input feature vector. The feature vector is standardized using zero-mean normalization to eliminate dimensional differences. The support vector machine classifier uses a radial basis function as the kernel function, with the kernel parameter set to 0.5 and the penalty parameter set to 100. The classifier optimizes the hyperparameters using a grid search method, with the search range being 0.001 to 10 for the kernel parameter and 1 to 1000 for the penalty parameter. The fault symptom identifier outputs the probability values ​​of various faults, including four types: blockage faults, void faults, fluidization non-uniformity faults, and abnormal heat transfer faults. When the probability value of any fault type exceeds 0.6, an abnormal state matrix is ​​established, and the matrix records the abnormal information according to time and spatial location.

[0043] The specific implementation of step S06 involves constructing a multi-source information fusion-based comprehensive assessment system for abnormal states. This step assesses the severity of detected abnormal states and confirms the fault type. The abnormal state assessment function employs a weighted fusion algorithm. Input parameters include five data sources: an abnormal state matrix, a pressure fluctuation matrix, a frequency domain gain matrix, historical fault data, and current operating condition parameters. Historical fault data is obtained by querying a fault database, containing information such as fault occurrence time, fault type, fault location, and handling measures. The database uses a relational database structure to store historical records. Current operating condition parameters are provided by the latest data from the operating condition parameter matrix and are used to assess the impact of current operating conditions on fault development. The assessment function uses fuzzy logic reasoning to establish a fuzzy rule base for fault severity. The rule base contains 100 reasoning rules covering combinations of different fault types and operating conditions. The fuzzification process uses a triangular membership function, the fuzzy reasoning uses the Mamdani reasoning method, and the defuzzification uses the centroid method to obtain numerical assessment results. The fault type classification results include four types: blockage faults, void faults, non-uniform fluidization faults, and abnormal heat transfer faults. Fault severity levels are divided into three grades: minor, moderate, and severe. Minor faults correspond to an evaluation value of 0 to 0.3, moderate faults to 0.3 to 0.7, and severe faults to 0.7 to 1.0. The evaluation function outputs the final fault determination result, including fault type, severity level, and recommended handling measures, providing decision support for operations and maintenance personnel.

[0044] The specific implementation of step S07 involves establishing a graded early warning and automatic response fault warning mechanism. This step aims to provide early warnings and trigger corresponding processing procedures before a fault occurs. The warning mechanism employs a multi-level threshold judgment strategy, setting three levels: warning threshold, alarm threshold, and emergency shutdown threshold. The warning threshold is set to a fault probability value of 0.3. When potential fault symptoms are detected, a yellow warning is activated, and the system automatically increases the monitoring frequency and records abnormal information. The alarm threshold is set to a fault probability value of 0.6. When fault symptoms are obvious, an orange alarm is activated, and the system automatically sends alarm information to the monitoring center and activates the emergency plan. The emergency shutdown threshold is set to a fault probability value of 0.8. When a serious fault is detected, a red emergency shutdown is activated, and the system automatically cuts off the feed and stops the fluidizing fan. The warning procedure adopts an event-driven architecture, using a message queue mechanism to achieve asynchronous processing, ensuring the real-time nature of the warning response. The monitoring system interface uses a graphical display method, showing the system operating status, fault probability trend, and warning level information in real time, with a refresh rate of 1Hz. The early warning information includes the fault type, location, severity, and handling suggestions, and is communicated to maintenance personnel via SMS, email, and audible and visual alarms. The system also features fault prediction capabilities, analyzing historical trends and current state changes to predict the likely time of a fault occurrence. The prediction algorithm employs a long short-term memory neural network, with a prediction time window of the next four hours.

[0045] like Figure 2 As shown, the detailed structure of the fluidized bed recognition model is based on a liquid neural network architecture. The input layer of this model contains five input nodes, corresponding to operating parameters such as bed temperature, fluidizing velocity, pressure drop, gas flow rate, and bed height. Each input node is configured with a normalization processing unit to eliminate dimensional differences between different parameters. The liquid computation layer is the core of the model, containing 200 dynamic neurons. The neurons' dynamic behavior is described using an integral firing model, and their membrane potentials are described by a first-order differential equation. Connections between neurons employ a sparse random connection method with a connection probability set to 20%. Connection weights are initialized with Gaussian random values, and weight updates use a backpropagation algorithm based on sparse connections. The temporal memory mechanism of the liquid computation layer is implemented through recursive connections. The output of each neuron is not only passed to the next layer but also fed back to other neurons in the current layer through delayed connections, forming a complex temporal dynamic network. The sparse attention mechanism dynamically adjusts connection weights according to the importance of the input signal. Attention weight calculation is based on the gradient information of the input signal, and an attention gating mechanism is used to selectively transmit information. The readout layer adopts a fully connected structure, containing 50 hidden nodes and 1 output node. The activation function of the hidden layer is the hyperbolic tangent function, and the activation function of the output layer is the sigmoid function to ensure that the output value is in the range of 0 to 1.

[0046] The detailed steps for establishing the training dataset for the fluidized bed recognition model include multi-stage data acquisition and processing. First, operational tests under different conditions were conducted on a fluidized bed test bench. Three typical operating conditions—normal fluidization, non-uniform fluidization, and abnormal fluidization—were artificially constructed by adjusting parameters such as fluidizing velocity, bed temperature, and particle loading. Normal fluidization corresponds to a fluidizing velocity ranging from 1.5 to 3 times the minimum fluidizing velocity, uniform bed temperature distribution, and a pressure drop stable within ±10% of the theoretically calculated value. Non-uniform fluidization is achieved by locally reducing the fluidizing airflow or increasing particle viscosity, resulting in varying degrees of fluidization in localized areas of the bed and fluctuating pressure drops. Abnormal fluidization is achieved by blocking some air inlets or adding large particulate impurities, resulting in severely non-uniform bed fluidization or complete loss of flow. More than 1000 sets of multi-dimensional time-series data were collected for each operating condition, with each set of data acquired for 300 seconds at a frequency of 10 Hz to ensure the sufficiency and representativeness of the data. In the data preprocessing stage, Kalman filtering was used to eliminate sensor noise, outliers were identified and removed using a 3-standard-deviation criterion, and missing data were filled using interpolation methods. A sliding window segmentation method was used to divide the continuous time-series data into fixed-length training samples, with a window length of 30 seconds corresponding to 300 data points and a window overlap rate of 50% to increase the number of training samples. Labeling was achieved through a combination of manual and automatic annotation, leveraging expert experience to determine the fluidization state category, and a supervised learning dataset containing input feature vectors and corresponding labels was established.

[0047] The detailed training steps of the fluid recognition model employ a phased optimization strategy to ensure model performance. The dataset is divided into training, validation, and testing datasets in a 7:2:1 ratio to ensure a balanced distribution of samples across different datasets. During model initialization, the Xavier initialization method is used to set network weights, with an initial learning rate of 0.001. An adaptive learning rate adjustment strategy dynamically adjusts the learning rate based on training progress. The training process uses the mini-batch gradient descent algorithm with a batch size of 32 and 200 training epochs. The cross-entropy loss function is used to measure the difference between predicted values ​​and true labels, and the optimizer uses an adaptive moment estimation algorithm for parameter updates. During training, model performance is evaluated using cross-validation, employing accuracy, precision, recall, and... The model's performance is comprehensively evaluated using metrics such as scores. The convergence condition is set to a change in the validation set loss function of less than 0.001 over 20 consecutive training epochs to avoid overfitting. Regularization techniques employ weight decay and dropout methods to prevent overfitting; the weight decay coefficient is set to 0.0001, and the dropout probability is set to 0.2. After training, the model's generalization ability is evaluated using a test set to ensure that the model's recognition accuracy exceeds 95% under different working conditions, meeting the performance requirements of practical applications.

[0048] Fluidized bed identification models can effectively handle the nonlinear dynamic characteristics and multi-scale temporal correlations of fluidized bed systems. Traditional fault detection methods mainly rely on threshold judgment of single parameters or simple statistical analysis, which cannot capture the complex dynamic behavior of fluidized bed systems and the coupling relationships between multiple parameters. Liquid neural networks, however, possess inherent temporal modeling and nonlinear mapping capabilities, enabling them to learn the temporal evolution and interaction mechanisms of parameters under various operating conditions during fluidized bed operation. Compared to traditional recurrent neural networks, liquid neural networks have the advantage of a sparse connection structure and dynamic weight adjustment mechanism, which can better handle long-term temporal dependencies and adapt to changes in system parameters. Compared to existing static classification methods such as support vector machines, liquid neural networks can learn and update model parameters in real time, exhibiting stronger adaptability and robustness. The model's sparse attention mechanism can automatically identify the most important feature parameters for fluidization state determination, avoiding the subjectivity and limitations of manual feature selection. The temporal memory mechanism can capture the historical trends of fluidized bed state changes, providing important temporal information for fault prediction. The dynamic weight adjustment mechanism allows the model to adapt to different operating conditions and particle characteristics, improving the model's versatility and practicality.

[0049] The key technical ideas of this invention are mainly reflected in the following aspects. The first key technical idea is a bed structure anomaly identification technology based on pressure pulse propagation detection. This technology achieves early identification of changes in the internal structure of the bed through an active excitation acoustic detection method. Compared with traditional passive monitoring methods, active excitation technology can detect local changes in the bed structure when fault symptoms are weak, exhibiting higher detection sensitivity and timeliness. The differences in the propagation characteristics of pressure pulses in media of different densities provide a physical basis for fault location. By analyzing changes in wave impedance, the location and severity of structural faults such as blockages and cavities can be accurately identified. The second key technical idea is a multi-scale frequency domain feature extraction technology based on wavelet transform. This technology decomposes the time-domain signal into components of different frequency bands, enabling accurate identification of fault symptoms at different scales. Traditional methods often only focus on time-domain features or single-frequency band features, failing to effectively distinguish between different types and severity of faults. Multi-scale frequency domain analysis can simultaneously capture changes in large-scale fluidization fluctuations, medium-scale particle motion, and small-scale turbulence characteristics of the bed, providing richer feature information for fault type determination. The third key technological approach is intelligent fluidized bed identification technology based on liquid neural networks. This technology automatically learns the complex mapping relationship between fluidized bed operating states and fault types through deep learning methods. Compared to traditional experience-based judgment or simple threshold methods, intelligent identification technology has stronger adaptability and identification accuracy, and can handle multi-parameter coupling and nonlinear dynamic characteristics. The temporal memory and attention mechanism of liquid neural networks enable the model to fully utilize historical information and important features, improving the accuracy and stability of fault identification.

[0050] The synergistic effect of these key technological approaches forms a complete multi-level fault detection system, offering significant technological advantages over existing technologies. Pressure pulse propagation detection technology provides a physical means of detecting bed structure anomalies; wavelet transform technology enables precise extraction of multi-scale features; and liquid neural network technology completes intelligent fault identification and state assessment. These three technological links complement each other and progress layer by layer, forming a complete technological chain from physical detection to feature extraction and intelligent identification. The multi-source information fusion mechanism organically integrates information obtained from different technologies, avoiding the limitations of single technologies and improving the reliability and accuracy of fault detection. The graded early warning mechanism provides corresponding handling strategies based on faults of varying severity, achieving full-process risk management from prevention to emergency response. The entire technological system boasts advantages such as high detection sensitivity, high identification accuracy, strong adaptability, and fast response speed, providing reliable technical support for the safe and stable operation of fluidized bed energy storage systems.

[0051] Specifically, the principle of this invention is as follows: The fundamental principle behind the technical solution of this invention in solving the problem of early identification of internal structural anomalies in fluidized bed energy storage systems lies in the establishment of an internal monitoring system for the bed based on acoustic propagation mechanisms. When structural anomalies such as particle aggregation or void formation occur inside the bed, changes in local medium density and elastic modulus directly affect the propagation speed and attenuation characteristics of pressure pulses. By calculating the changes in wave impedance values ​​at each monitoring point, the structural changes inside the bed can be accurately reflected. This invention employs a pressure sensor array arrangement strategy to achieve comprehensive monitoring of the pressure pulse propagation process at different heights within the bed. Combined with wavelet transform algorithms, frequency domain analysis of multi-dimensional parameters such as bed pressure drop gradient, particle concentration, and heat transfer coefficient is performed, enabling the extraction of fault symptom features in different frequency bands and effectively distinguishing between normal fluctuations and abnormal signals. The support vector machine fault symptom identifier establishes an accurate mapping model from multi-dimensional monitoring data to fault types by learning the complex nonlinear relationship between the energy distribution characteristics and wave impedance statistical characteristics of the main frequency bands. This achieves accurate classification and identification of different abnormal states such as blockage faults, void faults, and fluidization inhomogeneity faults. The fluidization identification model is based on a liquid neural network architecture. It captures the complex dynamic features of the bed fluidization process through dynamic weight adjustment and temporal memory mechanism, providing an accurate assessment of the overall fluidization state of the bed and providing important contextual information for fault determination. The entire technical solution constructs a complete technical chain from signal acquisition, feature extraction, pattern recognition to fault location through multi-level and multi-angle information fusion analysis.

[0052] 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.

[0053] The specific implementation of step S01 involves establishing a multi-sensor fusion operating condition parameter matrix, which organizes data according to two dimensions: time series and monitoring parameters. The mathematical expression for the operating condition parameter matrix is: ; In the formula, The working condition parameter matrix has dimensions of . ; For the first The bed temperature at any given moment, in °C, with a measurement range of 20 °C to 800 °C; For the first The fluidization velocity at any given moment, in m / s, with a measurement range from 0.1 m / s to 10 m / s; For the first The pressure drop at any given moment is measured in kPa, with a measurement range of 0 kPa to 50 kPa. For the first Gas flow rate at any given moment, in units of / h, measurement range is 0 / h to 1000 / h; For the first The bed height at any given time, in meters, with a measurement range of 0.5m to 3m; For the first Each sampling time is measured in seconds (s). This represents the total number of sampling points, with a default value of 3000 corresponding to a 300-second sampling duration. The data acquisition system uses a moving average filter to eliminate high-frequency noise. The expression for the filtered signal is: ; In the formula, For the first The filtered signal at each moment; For the first The raw sensor signal at each moment; The filter window length is set to 5 sampling points; This is the index variable within the window.

[0054] The specific implementation of step S02 is based on calculating the propagation characteristic parameters of the pressure pulse in the bed according to the acoustic wave propagation principle. The expression for calculating the signal propagation time difference between adjacent sensors using the cross-correlation algorithm is as follows: ; In the formula, It is a cross-correlation function; For upstream sensor signals; It is the complex conjugate of the downstream sensor signal; For time delay; The propagation time difference is a signal time variable. It is obtained by finding the peak value of the cross-correlation function. ; In the formula, The propagation time difference of the pressure pulse is expressed in seconds (s). The formula for calculating the propagation velocity of the pressure pulse is: ; In the formula, The propagation speed of the pressure pulse is expressed in m / s. The distance between adjacent sensors is set to 0.2m. The formula for calculating the wave impedance value is: ; In the formula, The acoustic impedance value is expressed in kg / (m²·s). The density of the mixed medium is expressed in kg / m³. The density of the mixed medium is calculated using a volume-weighted average method. ; In the formula, This refers to particle density, expressed in kg / m³, with a typical value of 2500 kg / m³. This refers to the density of a gas, expressed in kg / m³. The density of air at room temperature is approximately 1.2 kg / m³. The porosity of the bed is dimensionless and ranges from 0.4 to 0.8. The formula for calculating the rate of change of wave velocity is: ; In the formula, The wave velocity is the rate of change, which is dimensionless. The reference wave velocity is expressed in m / s, and is the average propagation velocity under normal fluidization conditions. The formula for calculating the rate of change of wave impedance is: ; In the formula, The rate of change of wave impedance is dimensionless. The historical average wave impedance is expressed in kg / (m²·s). The pressure pulse attenuation coefficient is obtained by fitting an exponential attenuation model, the expression of which is: ; In the formula, Distance The amplitude of the pressure pulse at the location, in Pa; The initial amplitude is expressed in Pa. This is the attenuation coefficient, in meters (m). - ¹; The propagation distance is in meters (m). The attenuation coefficient is obtained by least squares fitting, and the objective function is: ; In the formula, This represents the number of measurement points. For the first The actual amplitude at each measurement point; For the first The distance between measurement points. The expression for the pressure fluctuation matrix is: ; In the formula, This is a pressure fluctuation matrix with dimension 1. ; The total number of monitoring points is set to the bed height divided by 0.2m.

[0055] The specific implementation of step S03 involves using wavelet transform to perform multi-scale signal decomposition and calculate frequency domain features. The formula for calculating the bed pressure drop gradient is: ; In the formula, This represents the pressure drop gradient in the bed, expressed in kPa / m. and The pressure values ​​at adjacent pressure measuring points are in kPa. The distance between adjacent measuring points is set to 0.2m. The formula for calculating the heat transfer coefficient is: ; In the formula, The heat transfer coefficient is expressed in W / (m²·K). The heat flux density of the bed wall is expressed in W / m². This represents the temperature difference between the wall surface and the bed layer, expressed in Kelvin (K). The formula for calculating the bed layer porosity is: ; In the formula, The volume of the bed is in m³. This represents the total volume of solid particles, expressed in m³. The formula for calculating bed volume is: ; In the formula, This refers to the cross-sectional area of ​​the bed layer, in m². Here is the bed height, in meters (m). The formula for calculating the total volume of solid particles is: ; In the formula, The value represents the particle mass, expressed in kg. The mathematical expression for wavelet transform is: ; In the formula, These are wavelet coefficients, dimensionless; is a scale parameter, dimensionless, ranging from 1 to 64; These are translation parameters, with the same units as the input signal; For input signals; For the complex conjugate of the de Besi wavelet function; The signal time variable is used. The formula for calculating the energy distribution of wavelet coefficients is based on Passevar's theorem: ; In the formula, For the first Energy of layer wavelet coefficients; For the first Layer Wavelet coefficients; This refers to the number of coefficients. The formula for calculating the energy percentage of a frequency band is: ; In the formula, For the first The energy percentage of a frequency band, dimensionless; The total number of decomposition layers is set to 6. The expression for the frequency domain gain matrix is: ; In the formula, The frequency domain gain matrix has dimensions of . ; The energy percentage in the low-frequency range of 0.1Hz to 1Hz; The energy percentage in the mid-frequency band from 1Hz to 10Hz; The energy percentage in the high-frequency band from 10Hz to 50Hz; the number of rows in the matrix corresponds to four state characteristic parameters: bed pressure drop gradient, particle concentration, heat transfer coefficient, and bed porosity.

[0056] The specific implementation of step S04 is the same as described above, and will not be repeated in detail here.

[0057] The specific implementation of step S05 involves constructing a multi-feature fusion input feature vector and performing fault symptom identification. The expression for the input feature vector is: ; In the formula, The input feature vector is 7-dimensional; The mean wave impedance is expressed in kg / (m²·s). The standard deviation of wave impedance is expressed in kg / (m²·s). This is the maximum rate of change of wave impedance, which is dimensionless. This represents the minimum rate of change of wave impedance, which is dimensionless. The formula for calculating the mean wave impedance is: ; In the formula, This represents the total number of monitoring points. The formula for calculating the standard deviation of wave impedance is: .

[0058] The formula for eigenvector standardization is: ; In the formula, The standardized feature vectors are dimensionless. The mean of the eigenvectors; Let be the standard deviation of the feature vectors. The decision function expression for the support vector machine classifier is: ; In the formula, For classification decision results; For testing purposes; For the first Lagrange multipliers of the support vectors; For the first Labels of the support vectors; For radial basis kernel functions; For bias terms; The number of support vectors; For the first There are 1 support vector. The expression for the radial basis function kernel is: ; In the formula, The kernel function parameter is set to 0.5; Let be the Euclidean distance norm.

[0059] The specific implementation of step S06 involves constructing an anomaly state evaluation system based on fuzzy logic. The fuzzy membership function is represented by a triangular function: ; In the formula, is the membership function of the fuzzy set, which is dimensionless and ranges from 0 to 1; For input variables; , , The parameters of the triangle membership function are the left endpoint, peak point, and right endpoint, respectively. Fuzzy inference uses the Mamdani method, and the rule expression is: ; In the formula, For the first The membership degree of a rule; is the membership function of the predecessor; is the membership function for the consequent; For input variables; For output variables. Defuzzification is calculated using the centroid method: ; In the formula, For the deblurring result; For the output variable's first A discrete value; This represents the membership value of the synthesized output; Let be the number of discretization points. The comprehensive expression for the abnormal state evaluation function is: ; In the formula, This is the final fault assessment value, dimensionless, ranging from 0 to 1; This is the assessment value for pressure fluctuations; This is a frequency domain feature evaluation value; These are historical fault assessment values; This is the current operating condition assessment value; , , , For the weighting coefficients, satisfying The default values ​​are 0.3, 0.3, 0.2, and 0.2, respectively.

[0060] The specific implementation of step S07 is to establish a multi-level threshold determination early warning mechanism. The expression for the early warning threshold determination function is: ; In the formula, This is a warning level, dimensionless. This represents the fault probability value, dimensionless, ranging from 0 to 1; warning levels 0, 1, 2, and 3 correspond to normal, yellow warning, orange alarm, and red emergency shutdown, respectively. The fault prediction algorithm uses time series analysis, and the expression for the prediction function is: ; In the formula, The predicted probability of future failures; The current moment; The forecast time window is set to 4 hours. Set the historical data window length to 24 hours. This is the prediction function.

[0061] It needs to be explained that the operating condition parameter matrix The design principle is based on the matrix representation of multidimensional time-series data. By organizing the measurement values ​​of different sensors at different times in a unified format, the standardized storage and processing of multi-source heterogeneous data is realized. Compared with the traditional single-parameter monitoring method, this matrix structure can completely preserve the spatiotemporal evolution information of the system's operating state, providing a structured data foundation for subsequent intelligent analysis.

[0062] Cross-correlation function Based on the time delay estimation theory in signal processing, the signal propagation time is determined by calculating the correlation between two signals under different time delays. This method has stronger noise resistance and higher time resolution compared to simple peak detection. Formula for pressure pulse propagation speed Based on the fundamental physical laws of acoustic propagation, this active detection method reflects changes in medium properties by measuring the difference in propagation time of sound waves in different media. Compared with traditional static pressure measurement, this active detection method can generate obvious signal responses when there are slight changes in the bed structure, which significantly improves the sensitivity of fault detection and early warning capability. Wave impedance calculation formula Combining acoustic impedance theory and multiphase fluid dynamics principles, this method quantifies the degree to which a medium impedes the propagation of sound waves by multiplying sound velocity and density. When particle aggregation or voids occur inside the bed, local density changes can lead to significant changes in wave impedance. This method has higher spatial resolution and structural sensitivity compared to traditional temperature and pressure monitoring. Exponential decay model Based on the attenuation mechanism of sound waves in porous media, the internal structural anomalies of the bed can be identified by fitting the variation law of pressure pulse amplitude with propagation distance. Compared with the method that only focuses on propagation velocity, attenuation characteristic analysis can provide additional information on bed density distribution and pore structure. Wavelet transform formula Based on the theory of multi-resolution analysis, it is possible to decompose signals in both the time and frequency domains. Compared with the limitation of Fourier transform, which can only provide frequency domain information, wavelet transform can accurately locate the distribution characteristics of abnormal signals in time and frequency, providing rich feature information for the accurate identification of different types of faults. Feature vector The construction integrates frequency domain energy distribution and wave impedance statistical characteristics. By organically fusing the characteristic parameters obtained from different physical principles, a multi-dimensional fault characterization space is formed. Compared with the single-feature judgment method, multi-feature fusion can significantly improve the accuracy and robustness of fault identification and effectively avoid misjudgment caused by single sensor failure or environmental interference. Radial basis kernel function Based on the similarity measurement principle of Gaussian distribution, the similarity is evaluated by calculating the Euclidean distance between samples. This kernel function can map the linearly inseparable feature space to a high-dimensional space to achieve linear separation, and has a stronger nonlinear fitting ability compared to linear kernel functions. Fuzzy membership function Based on fuzzy set theory, triangular functions are used to describe the degree to which variables belong to fuzzy concepts. Compared with the traditional hard threshold judgment method, fuzzy logic can handle boundary fuzziness and uncertainty information, making the fault assessment results more consistent with the gradual characteristics in actual engineering. Centroid Defuzzing Formula Based on the expected value calculation principle in probability theory, the fuzzy inference results are converted into precise values ​​by weighted averaging. This method can fully consider the contribution of all possible output values ​​and has better continuity and stability compared to the maximum membership method. Multi-level early warning threshold function A tiered response mechanism was established. By setting different fault probability thresholds corresponding to different handling strategies, compared with the traditional method of a single alarm threshold, the tiered early warning mechanism can provide corresponding response measures according to the severity of the fault, realizing risk control throughout the entire process from preventive maintenance to emergency shutdown, and significantly improving the safety and economy of the system.

[0063] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: A 300MW fluidized bed energy storage system uses quartz sand as the thermal storage medium, and the bed cross-sectional area is 25... The bed height is designed to be 2.8m, and the normal operating temperature range is 500–650℃. To ensure the safe and stable operation of the system, the technical team decided to use the fault determination method of this invention to establish a comprehensive monitoring and diagnostic system.

[0064] In step S01, the technical team arranged 15 ceramic diaphragm differential pressure sensors vertically at 0.2m intervals inside the bed, forming a pressure sensor array. Each sensor has a measurement accuracy of ±0.1kPa, a measurement range of 0–50kPa, and can withstand operating temperatures up to 700℃. K-type thermocouple temperature sensors were installed at heights of 0.7m, 1.4m, and 2.1m on the sidewall of the bed. The sensor protective sleeves were made of 316L stainless steel, with a measurement accuracy of ±1℃. A turbine flow meter installed in the air inlet duct has a measurement range of 0.1–10m / s and a measurement accuracy of ±0.05m / s. The impeller is made of titanium alloy to ensure long-term stability. The orifice plate flow meter has an orifice ratio of 0.6 and a measurement range of 0–800. The measurement accuracy is ±2%. The ultrasonic level gauge on the bed sidewall operates at a frequency of 40kHz, with a measurement range of 0.5–3m and an accuracy of ±5mm. The data acquisition system synchronously acquires all sensor signals at a frequency of 10Hz, and uses a moving average filtering algorithm with a length of 5 sampling points to eliminate high-frequency noise. The operating parameter matrix is ​​organized in time series, with each row containing the real-time values ​​of 5 basic parameters.

[0065] In step S02, the technical team installed an electromagnetic vibrator at the center of the bed bottom. This vibrator drives a 50mm diameter magnetic piston to generate pressure pulses via an electromagnetic coil. The excitation signal uses a 1-100Hz sweep frequency, the vibration amplitude is set to 0.5-3mm, the pulse duration is 0.1s, and the interval is 5s. A pressure sensor array detects the pulse propagation process at a 1000Hz sampling frequency, and a cross-correlation algorithm is used to calculate the signal propagation time difference between adjacent sensors. Under normal fluidization conditions, the pressure pulse propagation velocity is approximately 95m / s, and the medium density is determined by a particle concentration of 420. With a gas density of 1.2 The volume-weighted average yields a total density of 210. The wave impedance value is calculated as the product of the propagation velocity and the medium density, and is approximately 19950 under normal conditions. The technical team set a threshold of 15% for the rate of change of wave impedance; points exceeding this value are marked as outliers. The attenuation coefficient is calculated using an exponential attenuation model. The fitting was used to obtain, where Let α be the initial amplitude, α be the attenuation coefficient, and x be the propagation distance.

[0066] In step S03, the technical team calculated the pressure drop gradient of the bed using the pressure difference between adjacent measuring points spaced 0.2m apart. Under normal conditions, the pressure drop gradient is approximately 2.1 kPa / m. A laser scattering optical concentration meter is installed in the observation window on the side wall of the bed, using a 632.8nm wavelength laser, a 90° scattering angle for detection, a measurement accuracy of ±5%, and a measurement range of 0–80°. The heat transfer coefficient is calculated as the ratio of the heat flux density at the bed wall to the temperature difference. The measurement accuracy of the thin-film heat flux meter is ±3%, and the heat transfer coefficient is approximately 185 during normal operation. The bed porosity is determined by the bed volume of 70. Total volume of solid particles 28 The difference, divided by the bed volume, yields 0.6. The four state characteristic parameters are decomposed into six levels using the de Bessie db4 wavelet to obtain wavelet coefficients for different frequency bands. The energy proportion of each frequency band is calculated using Passevar's theorem to construct the frequency domain gain matrix.

[0067] In step S04, the fluidization identification model constructed by the technical team adopts a liquid neural network architecture. The input layer receives 5-dimensional operating condition parameters. The liquid computation layer contains 200 dynamic neurons with a connection sparsity of 20%. The neuron connection weights are dynamically adjusted according to the bed temperature and fluidization velocity. The temporal memory length is set to 30 time steps, and the sparse attention mechanism adaptively adjusts the weights according to the pressure drop rate. The readout layer uses 50 fully connected neurons, and the output layer uses the sigmoid function. In actual operation, a bed fluidization state evaluation value of 0.25 is considered a normal fluidization state, an evaluation value of 0.58 is considered a non-uniform fluidization state requiring monitoring, and an evaluation value of 0.82 is considered an abnormal fluidization state requiring handling.

[0068] In step S05, the technical team extracted three features from the frequency domain gain matrix: the energy proportion in the low-frequency band (0.1–1 Hz), the energy proportion in the mid-frequency band (1–10 Hz), and the energy proportion in the high-frequency band (10–50 Hz). Under normal conditions, the low-frequency band accounts for 65%, the mid-frequency band for 28%, and the high-frequency band for 7%. The team also extracted wave impedance statistical features from the pressure fluctuation matrix; under normal conditions, the average wave impedance is 19950. Standard deviation 580 The maximum rate of change is 12%, and the minimum is -8%. The support vector machine uses radial basis functions, a kernel parameter of 0.5, and a penalty parameter of 100. An abnormal state matrix is ​​established when the probability of a blockage fault reaches 0.73.

[0069] During the implementation of step S06, the abnormal state evaluation function uses a weighted fusion algorithm to comprehensively analyze multi-source information. The technical team established a fuzzy rule base containing 100 inference rules, employing triangular membership functions and the Mamdani inference method. In a certain detection, the system identified a blockage fault type with a severity evaluation value of 0.45, classifying it as a medium fault. The evaluation results are shown in Table 1: Table 1. Results of Abnormal Status Assessment

[0070] Based on the comprehensive assessment in Table 1, the final fault determination result is a moderate blockage fault. It is recommended to increase the monitoring frequency and prepare to clean the equipment.

[0071] In step S07, the technical team established a three-level early warning mechanism: a yellow alert threshold of 0.3, an orange alert threshold of 0.6, and a red alert threshold of 0.8 for emergency shutdown. The real-time monitoring data detected by the system during a certain day's operation is shown in Table 2. Table 2 Real-time monitoring data records

[0072] According to the monitoring records in Table 2, the system progressed from a normal state to a serious malfunction within 15 minutes. The early warning system promptly detected the anomaly and took corresponding measures.

[0073] The technical team also established a historical fault database during implementation, recording fault information from the past 6 months. The statistical results are shown in Table 3. Table 3. Statistics on Historical Fault Types

[0074] Table 3 shows that blockage is the most common type of failure, occurring in 38.7% of cases, with an average processing time of 4.2 hours.

[0075] During the three months of actual operation, the fluidized bed energy storage system experienced 15 fault warnings, 12 of which were actual faults, resulting in a false alarm rate of 20%. The system's fault detection accuracy reached 93.2%, fault location accuracy was 91.8%, and the average fault detection time was 145 seconds. The fluidic state recognition model achieved an accuracy rate exceeding 95% under different operating conditions, meeting the requirements of practical applications.

[0076] It should be noted that the variables involved in this invention are explained in detail in Tables 4 and 5.

[0077] Table 4. Variable Explanation Table (Part 1)

[0078] Table 5. Variable Explanation Table (Part Two)

[0079] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for a fluidized bed energy storage system, characterized in that, Pressure sensor arrays are arranged at different heights within the fluidized bed energy storage system, while temperature and flow sensors are installed on the bed sidewalls to collect basic monitoring parameters and establish an operating parameter matrix. Periodic pressure pulses are generated by a vibration excitation source installed at the bottom of the bed. The pressure sensor array detects the propagation speed and attenuation characteristics of these pressure pulses in different bed regions, calculating the wave impedance values ​​at each monitoring point. Anomalies are marked when the wave impedance change rate exceeds a set threshold, establishing a pressure fluctuation matrix. State characteristic parameters are extracted and decomposed into multiple scales using a wavelet transform algorithm to obtain fault symptom signals in different frequency bands, establishing a frequency domain gain matrix. The operating parameters are then... A matrix-based fluidized bed identification model is used to analyze the fluidization state of the bed, and the model outputs an evaluation value of the fluidization state. A fault symptom identifier is established based on a support vector machine algorithm. The main frequency band energy distribution features are extracted from the frequency domain gain matrix, and the wave impedance statistical features are extracted from the pressure fluctuation matrix. The main frequency band energy distribution features and wave impedance statistical features are combined to form an input feature vector. The fault symptom identifier calculates the fault probability value. When the fault probability value exceeds a threshold, an abnormal state matrix is ​​established. An abnormal state evaluation function is used to comprehensively evaluate the abnormal state matrix. The abnormal state evaluation function outputs the final fault judgment result, and corresponding treatment measures are formulated based on the final fault judgment result. Establish a fault early warning mechanism. When potential fault signs are detected, the early warning procedure will be activated, and the final fault judgment result and handling suggestions will be output to the monitoring system interface.

2. The fault determination method for a fluidized bed energy storage system according to claim 1, characterized in that, The steps for establishing the operating condition parameter matrix are as follows: a two-dimensional data matrix is ​​formed by arranging the basic monitoring parameters of bed temperature, fluidizing velocity, pressure drop, gas flow rate, and bed height in a time series. The bed temperature is obtained by thermocouple temperature sensors installed at different heights of the bed, the fluidizing velocity is measured by a turbine flow meter installed in the air inlet duct, the pressure drop is measured by differential pressure sensors installed at the bottom and top of the bed, the gas flow rate is measured by an orifice plate flow meter in the air inlet duct, and the bed height is measured by an ultrasonic level gauge installed on the side wall of the bed.

3. The fault determination method for a fluidized bed energy storage system according to claim 2, characterized in that, The steps for establishing the pressure fluctuation matrix are as follows: Specifically, the pressure pulse propagation characteristic data matrix inside the bed is obtained through a pressure pulse propagation detection algorithm. This matrix includes the wave impedance value, wave velocity change rate, and attenuation coefficient parameters for each monitoring point. The wave impedance value is calculated by multiplying the pressure pulse propagation velocity and the medium density. The wave velocity change rate is calculated by the ratio of the difference between the current wave velocity and the reference wave velocity to the reference wave velocity. The attenuation coefficient is obtained by exponential fitting of the pressure pulse amplitude attenuation law.

4. The fault determination method for a fluidized bed energy storage system according to claim 3, characterized in that, The steps for establishing the frequency domain gain matrix are as follows: the frequency domain feature matrix is ​​obtained by performing multi-scale decomposition of the bed pressure drop gradient, particle concentration, heat transfer coefficient, and bed porosity state characteristic parameters using a wavelet transform algorithm. The bed pressure drop gradient is calculated by the ratio of the pressure difference to the distance between adjacent pressure measuring points. The particle concentration is obtained by online measurement using an optical concentration meter installed in the observation window on the bed sidewall. The heat transfer coefficient is calculated by the ratio of the bed wall heat flux density to the temperature difference. The bed porosity is calculated by the ratio of the difference between the bed volume and the total volume of solid particles to the bed volume.

5. The fault determination method for a fluidized bed energy storage system according to claim 4, characterized in that, The fluidization identification model refers to a bed fluidization state identification model based on a liquid neural network architecture, which includes three main parts: an input layer, a liquid computation layer, and a readout layer. The liquid computation layer adopts a sparsely connected dynamic neural network structure, and the connection weights between neurons are dynamically adjusted according to the bed temperature and fluidization velocity. The model captures the dynamic characteristics of the bed fluidization process through a temporal memory mechanism, and the attention weight parameters of the sparse attention mechanism are adaptively adjusted according to the pressure drop rate.

6. The fault determination method for a fluidized bed energy storage system according to claim 5, characterized in that, The steps for establishing the training dataset for the fluidized bed recognition model are as follows: collecting fluidized bed operation data under different working conditions, including three typical working conditions: normal fluidization state, non-uniform fluidization state, and abnormal fluidization state. More than 1,000 sets of multi-dimensional time-series data are collected for each working condition. The data acquisition frequency is 10Hz, and the single acquisition duration is 300s. Noise and outliers are eliminated through data preprocessing. The continuous time-series data is divided into training samples of fixed length using the sliding window method. Each training sample contains 30s of time-series data.

7. The fault determination method for a fluidized bed energy storage system according to claim 6, characterized in that, The training steps for the fluid recognition model specifically involve dividing the dataset into 70% for training, 20% for validation, and 10% for testing. An adaptive learning rate optimization algorithm is used to iteratively update the model parameters. During training, the model performance is evaluated using cross-validation. The convergence condition is set to a change in the validation set loss function of less than 0.001 within 20 consecutive training cycles. After training, the generalization ability of the model is evaluated using the test set to ensure that the recognition accuracy of the model exceeds 95% under different working conditions.

8. The fault determination method for a fluidized bed energy storage system according to claim 7, characterized in that, A fluidization state assessment value between 0 and 0.35 is considered normal fluidization; a value between 0.35 and 0.75 is considered non-uniform fluidization requiring further monitoring; and a value between 0.75 and 1 is considered abnormal fluidization.

9. The fault determination method for a fluidized bed energy storage system according to claim 8, characterized in that, The steps for establishing the training dataset for the fault symptom identifier specifically involve collecting fluidized bed operation data under different fault types and normal conditions. This includes five typical conditions: blockage fault, void fault, non-uniform fluidization fault, abnormal heat transfer fault, and normal operation. More than 500 sets of multi-dimensional monitoring data are collected for each condition, with a data acquisition duration of 600 seconds per set and an acquisition frequency of 5 Hz. The main frequency band energy distribution characteristics and wave impedance statistical characteristics are extracted from each set of data to form a feature vector.

10. The fault determination method for a fluidized bed energy storage system according to claim 9, characterized in that, The main frequency band energy distribution characteristics include three components: the energy proportion of the low frequency band (0.1Hz to 1Hz), the energy proportion of the mid frequency band (1Hz to 10Hz), and the energy proportion of the high frequency band (10Hz to 50Hz). The wave impedance statistical characteristics include four components: the wave impedance mean, the wave impedance standard deviation, the maximum wave impedance change rate, and the minimum wave impedance change rate. The seven-dimensional feature vector is combined with the corresponding fault type label to establish a supervised learning training dataset. The differences in the dimensions of different features are eliminated through data standardization.

Images