A method of flow battery system fault diagnosis

Abstract

The application discloses a kind of liquid flow battery system fault diagnosis method, belong to liquid flow battery system monitoring technical field, including the following steps: S1, carries out multidimensional parameter monitoring;S2, carries out dynamic data preprocessing;S3, establishes health state evaluation model;S4, carries out fault diagnosis and classification;S5, carries out self-adapting optimization mechanism.The application establishes electrolyte flow rate-pile polarization voltage multivariate regression model, and integrates vibration spectrum, spectrum analysis and other multidimensional characteristics, constructs high-dimensional fault feature space.Make the accuracy rate of distinguishing mass transfer efficiency anomaly and electrode failure improve, misjudgment rate reduces.Fault coverage rate rises to, support the accurate classification of multiple kinds of faults.

Description

Technical Field

[0001] This invention belongs to the field of flow battery system monitoring technology, and specifically relates to a method for fault diagnosis of flow battery systems. Background Technology

[0002] 1. Single parameter threshold monitoring and static analysis: Existing technologies generally rely on fixed threshold alarm mechanisms for isolated electrochemical parameters such as voltage and temperature (e.g., triggering shutdown when the voltage deviates from the rated value by ±5 V), which is essentially a static, single-point monitoring paradigm. Such methods can only capture local state changes, neglecting the complex multiphysics coupling effects within the electrochemical system. First, key fluid behaviors such as electrolyte flow rate, pressure distribution, and flow stability, which are crucial fluid dynamic parameters, were not included in the monitoring system. Abnormal flow rates may lead to increased concentration polarization at the electrode surface, while pressure fluctuations reflect the risk of flow channel blockage. Both directly affect reaction uniformity and thermal management efficiency.

[0003] Subsequently, dynamic information such as mechanical state parameters, including pump body vibration spectrum, bearing wear characteristics, and mechanical seal integrity, was not correlated and analyzed. For example, impeller imbalance in a centrifugal pump can trigger vibrations in a specific frequency band, which in turn induces electrolyte pulsation, ultimately leading to voltage fluctuations—but current technology cannot establish such cross-domain causal relationships.

[0004] 2. Offline alignment method based on feature curves: In fault detection systems for vanadium redox flow batteries, offline comparison based on characteristic curves is a commonly used strategy. The core of this method lies in pre-defining a set of standardized charge-discharge characteristic curves (such as a typical voltage-time curve) as a benchmark for battery health. During actual operation, the battery's charge-discharge data is collected, and its voltage-time trajectory is compared offline with the aforementioned pre-defined standard curves. This comparison primarily focuses on identifying whether voltage fluctuations or deviations exceed preset safety thresholds, thus serving as a preliminary basis for determining whether the battery exhibits abnormalities or potential faults (such as short circuits, internal leakage, or electrolyte imbalance).

[0005] However, this method has a significant limitation: it primarily relies on static or quasi-static comparisons of macroscopic electrochemical parameters such as voltage, failing to establish and integrate a dynamic coupling model between key electrochemical parameters (e.g., voltage, current density, state of charge (SOC)) and core fluid dynamics parameters (e.g., electrolyte flow rate, flow volume, and channel pressure drop). Specifically, this model should be able to quantitatively describe how changes in fluid state (e.g., a decrease in electrolyte flow rate due to pump failure or pipeline blockage) dynamically and in real time affect the electrochemical processes inside the battery (especially concentration polarization), leading to significant changes in polarization voltage. This lack of a dynamic coupling mechanism means that when diagnosing complex faults caused by fluid system anomalies (e.g., insufficient flow rate) that manifest as voltage anomalies, this method may fail to accurately distinguish the root cause of the fault or lead to misjudgments of voltage fluctuations, thus reducing the accuracy of diagnosis and the ability to warn of potential risks.

[0006] 3. Data acquisition and processing in a fixed window: The fixed data acquisition window causes a response delay of ≥3 seconds for high-speed faults (such as sudden pressure drops), and shortening the window will increase the false alarm rate due to noise interference.

[0007] 4. Lack of modeling for cross-dimensional dynamic coupling relationships: For example, the lack of a quantitative relationship between electrolyte flow rate and stack polarization voltage led to the misdiagnosis of decreased mass transfer efficiency as electrode failure, resulting in a high misjudgment rate.

[0008] 5. Insufficient ability to analyze complex faults: For example, the recognition rate for complex faults such as "diaphragm damage + electrolyte contamination" is low, and the primary and secondary relationships of the faults cannot be distinguished.

[0009] 6. The contradiction between real-time performance and diagnostic accuracy: Fixed data windows cause delays in high-speed fault response; for example, detecting a sudden pressure drop fault takes more than 3 seconds. Shortening the window introduces high-frequency noise, increasing the false alarm rate.

[0010] 3. Fixed-window data acquisition mechanism causes response delay. The existing system uses a fixed-duration data acquisition window (e.g., 3 seconds), resulting in an inherent delay in response to high-speed transient faults. When the system encounters millisecond-level faults such as sudden pressure drops, the fixed window requires waiting for a full cycle (≥3 seconds) to complete data acquisition and processing, causing a lag in fault response. If attempts are made to shorten the window duration to improve timeliness (e.g., reduce it to 1 second), the signal-to-noise ratio deteriorates significantly due to insufficient sampling time. High-frequency noise interference is abnormally amplified, triggering false alarms and increasing the false alarm rate beyond acceptable thresholds.

[0011] 4. Lack of modeling for cross-dimensional dynamic coupling relationships. The causal relationships between key parameters have not been quantitatively characterized, especially lacking a dynamic coupling model between electrolyte flow rate and stack polarization voltage. When abnormal flow rate leads to increased concentration polarization, the system, unable to identify the quantitative transfer function of "flow rate-voltage," mistakenly attributes voltage fluctuations, a characteristic of decreased mass transfer efficiency, to electrode activity decay or catalyst failure. This cross-domain mechanism confusion results in a diagnostic error rate of over 40% for the same phenomenon, and subsequent maintenance strategies deviate significantly from the actual fault source.

[0012] 5. Insufficient ability to collaboratively analyze complex faults. The system lacks sufficient accuracy in identifying coupled faults, particularly in decoupling the interactive effects of multiple faults. For example, in the scenario of "micro-damage to the diaphragm accompanied by electrolyte contamination," the system can only identify the presence of the fault (accuracy <65%), but cannot analyze: ① which is the primary cause of the voltage drop: ion channel leakage caused by diaphragm damage or electrolyte conductivity decrease due to impurity deposition; ② the temporal coupling strength between the two faults (e.g., whether impurities diffuse through the damaged diaphragm); ③ the quantitative ranking of the fault contribution (e.g., 70% due to damage vs. 30% due to contamination). This deficiency directly leads to insufficient targeting of repair solutions.

[0013] 6. The dilemma of the trade-off between real-time performance and diagnostic accuracy Fixed-window architectures struggle to balance high-speed fault detection with interference resistance. For example, in pressure drop fault detection, a 3-second window can guarantee a noise suppression rate of >95% (false alarm rate <5%), but the fault response delay is ≥3 seconds. When the window is compressed to 0.5 seconds, the delay decreases to 0.8 seconds, but high-frequency noise interference causes the false alarm rate to soar to >25%. Improving real-time performance requires reducing the window size, but a shorter window widens the system bandwidth into the noise-sensitive region, significantly increasing the false alarm rate. Summary of the Invention

[0014] To address the aforementioned problems, this invention proposes a fault diagnosis method for a flow battery system, comprising the following steps: S1. Perform multi-dimensional parameter monitoring; S2. Perform dynamic data preprocessing; S3. Establish a health status assessment model; S4. Perform fault diagnosis and classification; S5. Implement an adaptive optimization mechanism.

[0015] Furthermore, in step S1, electrochemical parameters such as stack voltage, current, temperature, and voltage balance data of individual cells are collected in real time.

[0016] Furthermore, in step S1, the electrolyte characteristics are monitored by flow sensors and pressure sensors to detect electrolyte flow rate and pressure; and the electrolyte concentration and impurity content are detected by spectral analysis or conductivity sensors.

[0017] Furthermore, in step S1, mechanical status parameters are used to identify blockages or leaks by monitoring the vibration signal of the circulating pump and changes in pipeline pressure.

[0018] Furthermore, step S2 includes the following steps: S21. Noise filtering; S22, Feature Extraction.

[0019] Furthermore, in step S21, noise filtering employs a dynamic multimodal filtering mechanism, selecting wavelet transform, Kalman filtering, or adaptive sliding window algorithm for signal denoising based on the characteristics of environmental noise.

[0020] Furthermore, in step S22, time-domain, frequency-domain, and time-frequency joint features are extracted from the voltage sequence.

[0021] Furthermore, in step S3, the stack performance index SPI evaluates the overall state of the stack by assessing the voltage uniformity of individual cells and the rate of change of internal resistance.

[0022] Furthermore, in step S4, a multi-level fault tree is analyzed: Level 1 diagnosis: based on threshold triggering; Secondary diagnosis: An improved convolutional neural network (CNN) is used to fuse features from multiple data sources to identify the fault type; Level 3 diagnosis: Combining expert systems and fuzzy logic, probabilistic reasoning is used to diagnose complex faults.

[0023] Further, step S5 includes: Dynamic window adjustment: Automatically adjusts the length of the data acquisition window based on the battery status to balance real-time performance and accuracy; Online model update: Incremental learning algorithm is introduced to update the diagnostic model parameters using new fault samples, thereby improving generalization ability.

[0024] The beneficial effects of this invention are as follows: Breakthrough in technical performance: This invention constructs a dynamic coupling model of electrolyte flow rate and stack polarization voltage, integrating multi-source data such as vibration spectrum and spectral analysis to achieve cross-dimensional fault feature fusion. This completely solves the problem of high false positive rates (>40%) for fluid anomalies and electrode failures in the background technology, improving the fault classification accuracy to >95%. Based on a three-level progressive diagnostic architecture (threshold warning / CNN classification / fuzzy inference), the composite fault recognition rate is improved from <65% to ≥92%. At the same time, the response speed for millisecond-level faults such as pressure drops is reduced from ≥3 seconds to 0.5 seconds, and the false alarm rate is stably controlled at <5%, breaking through the dilemma of mutual exclusion between real-time performance and diagnostic accuracy in the existing technology.

[0025] System adaptability and engineering value optimization: The dynamic window adjustment mechanism adaptively matches the data acquisition strategy according to the charging and discharging state (1.5 seconds steady-state window / 0.3 seconds transient window), improving the signal-to-noise ratio by >40% and reducing high-frequency noise interference to 1 / 3 of the original solution. The incremental learning algorithm updates the model parameters every 100 new samples received, narrowing the diagnostic accuracy fluctuation across battery types (such as all-vanadium and zinc-bromine flow batteries) from ±15% to ±3%. Combined with composite fault probability fusion technology, the maintenance solution is more targeted by 50%, maintenance costs are reduced by 30%, and the accident rate is reduced by >60%, significantly enhancing the long-term operational reliability of the system in scenarios such as new energy power plants.

[0026] Cross-dimensional dynamic coupling modeling and multimodal data fusion: A multivariate regression model of electrolyte flow rate and stack polarization voltage is established, and multidimensional features such as vibration spectrum and spectral analysis are integrated to construct a high-dimensional fault feature space. This improves the accuracy of distinguishing between abnormal mass transfer efficiency and electrode failure, and reduces the false positive rate. Fault coverage is increased, supporting accurate classification of various types of faults.

[0027] A three-tiered progressive diagnostic architecture optimizes response efficiency: It establishes a multi-level fault tree analysis, resulting in shorter time for resolving complex faults and improved identification rate. False alarm rate is reduced.

[0028] Adaptive optimization mechanisms enhance robustness and generalization: Dynamic window adjustment: The data acquisition window is automatically adjusted according to operating conditions, improving the signal-to-noise ratio and response speed; Incremental learning: The weights of the fully connected CNN layer are updated every 100 new samples received. Cross-battery type adaptability is enhanced, and accuracy fluctuations are reduced when switching between different types of flow battery models.

[0029] Composite Fault Probability Fusion and Evidence Theory Optimization: By fusing CNN output probabilities and expert rule confidence scores, the joint probability of composite faults is calculated. This improves the diagnostic accuracy of composite faults and reduces the false alarm rate. Attached Figure Description

[0030] Figure 1 Schematic diagram of flow battery system structure and data acquisition at monitoring points; Figure 2 Multi-level fault diagnosis flowchart; Figure 3 : Architecture diagram of a multi-source data fusion model based on CNN. Detailed Implementation

[0031] To make the technical means and objectives of this invention easier to understand, the invention is further described below with reference to specific embodiments. A fault diagnosis method for a flow battery system includes the following steps: S1. Multi-dimensional parameter monitoring Electrochemical parameters: Real-time acquisition of stack voltage, current, temperature, and voltage balance data of individual cells.

[0032] Electrolyte characteristics: Electrolyte flow rate and pressure are monitored by flow sensors and pressure sensors; electrolyte concentration and impurity content are detected by spectral analysis or conductivity sensors.

[0033] Mechanical status parameters: Monitor the vibration signal of the circulating pump and changes in pipeline pressure to identify blockages or leaks.

[0034] S2. Dynamic Data Preprocessing Noise Removal: A dynamic multimodal filtering mechanism is adopted, and wavelet transform, Kalman filtering or adaptive sliding window algorithm is selected for signal denoising based on the characteristics of environmental noise (such as electromagnetic interference and mechanical vibration).

[0035] Feature extraction: Extracting time-domain (e.g., volatility), frequency-domain (e.g., harmonic components), and time-frequency joint features (e.g., wavelet packet energy entropy) from voltage sequences.

[0036] S3. Health Status Assessment Model The Battery Stack Performance Index (SPI) assesses the overall condition of the battery stack by evaluating the voltage uniformity of individual cells and the rate of change of internal resistance.

[0037] S4. Fault Diagnosis and Classification Multi-level fault tree analysis: Level 1 diagnosis: based on threshold triggers (such as a sudden drop in electrolyte pressure triggering a "leakage" warning).

[0038] Secondary diagnosis: An improved convolutional neural network (CNN) is used to fuse features from multiple sources to identify fault types (such as electrode failure, diaphragm damage, etc.).

[0039] Level 3 diagnosis: Combining expert systems and fuzzy logic, probabilistic reasoning is used to solve complex faults (such as "pump blockage + electrolyte contamination").

[0040] S5. Adaptive Optimization Mechanism Dynamic window adjustment: The data acquisition window length is automatically adjusted according to the battery's operating status (such as charging and discharging stages) to balance real-time performance and accuracy.

[0041] Online model update: Incremental learning algorithm is introduced to update the diagnostic model parameters using new fault samples, thereby improving generalization ability.

[0042] Example 2 This invention discloses a fault diagnosis method for flow battery systems. By integrating multi-source data from electrochemical, mechanical, and fluid dynamics sources, and combining it with intelligent algorithms, it achieves early warning and precise fault location. This method can significantly improve the reliability and operation and maintenance efficiency of flow battery systems, and is applicable to scenarios such as new energy power plants and grid peak shaving, specifically including the following: Diagnosis of electrolyte leakage in vanadium redox flow batteries: 1. Multimodal data acquisition and dynamic preprocessing: Electrochemical parameters: The total voltage of the fuel cell stack and the voltage of each individual cell are monitored in real time using a distributed voltage acquisition module. The surface temperature and internal resistance of the fuel cell stack are also acquired simultaneously.

[0043] Fluid dynamics parameters: Electromagnetic flow meters and pressure transmitters are installed at key nodes in the electrolyte circulation pipeline (pile inlet, tank outlet), and the data is synchronously transmitted to the edge computing node.

[0044] Mechanical condition parameters: A triaxial vibration sensor is installed on the casing of the circulating pump to collect the RMS value and spectral characteristics of the vibration signal, and a Kalman filter is used to eliminate high-frequency mechanical noise.

[0045] 2. Dynamic window optimization and feature fusion: Adaptive window adjustment: During the charge / discharge switching phase (such as constant current to constant voltage), the data window is shortened to capture the characteristics of sudden pressure drops; during the steady-state phase, the window is extended to extract the low-frequency energy ratio of the pressure waveform.

[0046] Multi-source feature fusion: Constructing multi-dimensional feature vectors, including: Time-domain characteristics: pressure change rate, voltage fluctuation rate, vibration RMS value; Frequency domain characteristics: energy proportion of pressure signal in leakage characteristic frequency band, and voltage distortion rate of fuel cell stack.

[0047] 3. The execution process of the three-level diagnostic architecture, such as... Figure 2 As shown: Level 1 warning: When the pressure drops suddenly and the voltage fluctuation rate is high, a suspected leakage alarm is triggered.

[0048] Second-level CNN classification: Input the stress time series signal into the improved CNN model (structure as follows) Figure 3 The model consists of three convolutional kernels, a max-pooling layer, and a feature fusion layer. The training data includes various leakage scenarios, with a sample size of ≥2000 sets, and training continues until convergence.

[0049] Three-level expert system reasoning: Fuzzy rule: If the pressure drop rate exceeds the threshold and the vibration spectrum shows increased energy in the leakage characteristic frequency band, then the leakage confidence increases; Evidence fusion: Calculate the composite probability, and locate the leak point when the leakage confidence exceeds the threshold.

[0050] 4. Adaptive response strategy: After a Level 1 warning is triggered, the system automatically switches to the backup pipeline and limits the output power of the fuel cell stack.

[0051] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for a flow battery system, characterized in that, Includes the following steps: S1. Perform multi-dimensional parameter monitoring; S2. Perform dynamic data preprocessing; S3. Establish a health status assessment model; S4. Perform fault diagnosis and classification; S5. Implement an adaptive optimization mechanism.

2. The fault diagnosis method for a flow battery system as described in claim 1, characterized in that, In step S1, electrochemical parameters such as stack voltage, current, temperature, and voltage balance data of individual cells are collected in real time.

3. The fault diagnosis method for a flow battery system as described in claim 1, characterized in that, In step S1, the electrolyte characteristics are monitored by flow sensor and pressure sensor to detect electrolyte flow rate and pressure; electrolyte concentration and impurity content are detected by spectral analysis or conductivity sensor.

4. The fault diagnosis method for a flow battery system as described in claim 1, characterized in that, In step S1, mechanical condition parameters are monitored by observing the vibration signal of the circulating pump and changes in pipeline pressure to identify blockages or leaks.

5. The fault diagnosis method for a flow battery system as described in claim 1, characterized in that, Step S2 includes the following steps: S21. Noise filtering; S22, Feature Extraction.

6. The fault diagnosis method for a flow battery system as described in claim 5, characterized in that, In step S21, noise filtering adopts a dynamic multimodal filtering mechanism, and selects wavelet transform, Kalman filtering or adaptive sliding window algorithm to denoise the signal according to the characteristics of environmental noise.

7. The fault diagnosis method for a flow battery system as described in claim 5, characterized in that, In step S22, time-domain, frequency-domain, and time-frequency joint features are extracted from the voltage sequence.

8. The fault diagnosis method for a flow battery system as described in claim 1, characterized in that, In step S3, the stack performance index SPI evaluates the overall state of the stack by assessing the voltage uniformity of individual cells and the rate of change of internal resistance.

9. The fault diagnosis method for a flow battery system as described in claim 1, characterized in that, In step S4, the multi-level fault tree is analyzed: Level 1 diagnosis: based on threshold triggering; Secondary diagnosis: An improved convolutional neural network (CNN) is used to fuse features from multiple data sources to identify the fault type; Level 3 diagnosis: Combining expert systems and fuzzy logic, probabilistic reasoning is used to diagnose complex faults.

10. The fault diagnosis method for a flow battery system as described in claim 1, characterized in that, Step S5 includes: Dynamic window adjustment: Automatically adjusts the length of the data acquisition window based on the battery status to balance real-time performance and accuracy; Online model update: Incremental learning algorithm is introduced to update the diagnostic model parameters using new fault samples, thereby improving generalization ability.

Images