A Deep Learning-Based Method and System for SOFC State Monitoring and Health Diagnosis

Abstract

This invention provides a deep learning-based method and system for SOFC state monitoring and health diagnosis, relating to the field of solid oxide battery state monitoring technology. The method includes: collecting SOFC operational data; preprocessing the operational data and extracting features from the preprocessed data; inputting the operational data into a deep learning diagnostic model, which employs a hybrid architecture of convolutional neural networks and long short-term memory networks, combined with an attention mechanism to output an adaptive health status index (HSI); and real-time monitoring based on the HSI and its rate of change, issuing an early warning if the HSI falls below a preset threshold or the HSI rate of change exceeds a preset threshold. This invention effectively solves key problems in SOFC state monitoring and health diagnosis, such as data complexity, insufficient accuracy, and poor real-time performance, providing an innovative technical solution for efficient SOFC operation and long-life management.

Description

Technical Field

[0001] This invention relates to the field of solid oxide battery state monitoring technology, specifically to a deep learning-based method and system for SOFC state monitoring and health diagnosis. Background Technology

[0002] Solid oxide fuel cells (SOFCs) are a highly efficient, low-emission energy conversion technology. With their high energy efficiency and fuel adaptability, SOFCs have significant potential in distributed power generation, aerospace, and industrial energy systems.

[0003] However, in existing related technologies, the condition monitoring and health diagnosis of SOFCs still have the following problems:

[0004] First, the data is complex and feature extraction is difficult. The data generated by SOFC operation is high-dimensional, time-varying, and strongly coupled, making it difficult for traditional physical models and shallow machine learning methods to effectively extract its spatiotemporal features.

[0005] Second, the diagnostic accuracy and robustness are insufficient. Existing technologies are prone to false alarms or missed alarms when multiple faults coexist or when operating conditions are complex.

[0006] Third, the real-time performance is insufficient. Faults in SOFC operation often develop rapidly, and traditional monitoring systems have long response times, making it difficult to achieve early warning.

[0007] Fourth, there is a lack of full lifecycle management. Existing technologies are mostly focused on a single stage (such as fault detection), lacking a closed-loop system from data acquisition to diagnosis and then to optimization feedback, which makes it impossible to achieve intelligent and continuous health management of SOFCs. Summary of the Invention

[0008] In order to at least partially solve the technical problems in the related technologies, the present invention provides a method and system for SOFC state monitoring and health diagnosis based on deep learning.

[0009] To achieve the above objectives, the technical solution adopted by the present invention includes:

[0010] According to a first aspect of the present invention, a method for SOFC state monitoring and health diagnosis based on deep learning is provided, comprising the following steps:

[0011] Step S1: Collect SOFC operating data, including temperature, total stack voltage, operating current, hydrogen flow rate, oxygen flow rate, and SOFC internal pressure;

[0012] Step S2: Preprocess the collected operating data and extract features from the preprocessed operating data. The features include voltage attenuation rate, voltage mean, voltage standard deviation, and voltage peak value.

[0013] Step S3: Input the preprocessed running data into the deep learning diagnostic model. The deep learning diagnostic model adopts a hybrid architecture of convolutional neural network and long short-term memory network, and combines an attention mechanism to output an adaptive health status index. ;

[0014] in, The calculation process specifically includes:

[0015] Step S3-1: Calculate the basic adaptive health status index :

[0016]

[0017] Step S3-2: Calculate the rate of change:

[0018]

[0019] Step S3-3: Adjust the weights:

[0020]

[0021] Step S3-4: Calculation :

[0022]

[0023] Where, First, set the weights. Set a weight for the second one. Voltage attenuation rate, For SOFC operating temperature, For the optimal operating temperature of SOFC, The maximum allowable operating temperature for SOFC. The adaptive health status index of the previous moment is used to calculate the rate of change and reflect the temporal correlation of the state. Adjust the weight first. The second adjustment weight;

[0024] Step S4: According to Real-time monitoring of its rate of change, if Below the preset threshold or If the rate of change exceeds a preset threshold, an early warning will be issued.

[0025] Optionally, the deep learning-based SOFC state monitoring and health diagnosis method further includes:

[0026] Step S5: In response to the warning results, display the health status of SOFC through a visual interface and provide optimization suggestions.

[0027] Optionally, step S5 specifically includes:

[0028] In response to the warning results, the health status of SOFC is displayed through a visual interface;

[0029] Output the corresponding optimization suggestions based on the following formula:

[0030]

[0031] In the formula, To optimize the suggestions, This refers to the actual operating temperature of the SOFC.

[0032] Optionally, step S2 specifically includes:

[0033] Step S2-1: Denoise the collected running data based on wavelet transform;

[0034] Step S2-2: Normalize the noise-reduced running data;

[0035] Step S2-3: Extract features from the normalized running data.

[0036] Optionally, the deep learning diagnostic model includes a CNN part and an LSTM part;

[0037] The CNN part includes two convolutional layers, each followed by a ReLU activation function, and then dimensionality reduction is performed through a max pooling layer after the convolution operation.

[0038] The LSTM part includes two LSTM layers, each followed by a Dropout layer to prevent overfitting.

[0039] Optionally, the attention mechanism is implemented in the following way:

[0040] Step S3-2-1: Calculate the feature vector score output by the Long Short-Term Memory network;

[0041] Step S3-2-2: Normalize the feature vector scores;

[0042] Step S3-2-3: Weight key features to highlight their importance. The contribution of the assessment.

[0043] Optionally, step S4 specifically includes:

[0044] Calculate according to the following formula Decline rate:

[0045]

[0046] Where, Let be the adaptive health status index at time t - k. For time intervals;

[0047] when When the value is greater than the first preset threshold, the first warning signal is output;

[0048] when When the value is less than the second preset threshold, a second warning signal is output.

[0049] Optionally, the The value is 5, and the first preset threshold is set to 0.05; and / or,

[0050] The second preset threshold is set to 0.7.

[0051] According to a second aspect of the present invention, a deep learning-based SOFC state monitoring and health diagnosis system is also provided, for executing the deep learning-based SOFC state monitoring and health diagnosis method described in any of the technical solutions of the first aspect of the present invention, wherein the deep learning-based SOFC state monitoring and health diagnosis system comprises:

[0052] The data acquisition module is used to collect SOFC's runtime data;

[0053] The preprocessing module is used to preprocess the collected running data and extract features from the preprocessed running data.

[0054] The deep learning diagnostic model employs a hybrid architecture of convolutional neural networks and long short-term memory networks, combined with an attention mechanism. This model is used to analyze data and output an adaptive health status index. ;

[0055] The status monitoring and early warning module is used to... It monitors and issues early warnings in real time about its rate of change.

[0056] Optionally, the deep learning-based SOFC state monitoring and health diagnosis system further includes:

[0057] The visualization and feedback module is used to respond to the warning results, display the health status of SOFC through a visual interface, and provide optimization suggestions.

[0058] Beneficial effects:

[0059] 1. Through the above technical solution, firstly, this invention is the first to combine Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM) to design a novel deep learning model for multidimensional temporal series data of SOFC. CNN is used to extract spatial features, while LSTM is used to capture temporal dependencies. Compared to the existing HY-CNN-NARX paradigm and intelligent diagnostic methods, the hybrid model of this invention is superior in spatiotemporal feature analysis and can achieve higher accuracy in health status assessment.

[0060] Second, the adaptive health status index proposed in this invention The algorithm, by dynamically adjusting parameter weights, can quantify the health status of SOFC in real time. Compared with existing multi-label diagnostic methods, this algorithm can self-optimize according to changes in operating conditions, effectively improving the flexibility and accuracy of diagnosis.

[0061] Third, the method of this invention integrates data acquisition, preprocessing, diagnosis, and early warning steps, enabling the effective and accurate construction of a closed-loop management system for the entire SOFC lifecycle. Compared with traditional diagnostic methods and cloud platform solutions in existing related technologies, the closed-loop design of this invention can effectively reduce manual intervention and significantly improve the system's intelligence level and operational continuity.

[0062] Fourth, the method of this invention, through data augmentation techniques and attention mechanisms, can effectively improve the robustness of the model in complex environments. Simultaneously, by combining algorithm optimization and hardware acceleration, this invention can achieve rapid response (e.g., within 0.5 seconds) and a fault prediction window of several hours, far exceeding the real-time performance of existing related technologies.

[0063] In summary, this invention introduces a deep learning diagnostic model and an adaptive health status index. The closed-loop system covering the entire lifecycle can effectively solve key problems such as data complexity, insufficient accuracy, and poor real-time performance in SOFC condition monitoring and health diagnosis, providing an innovative technical solution for the efficient operation and long-life management of SOFC.

[0064] 2. Other beneficial effects or advantages of the present invention will be described in detail in the specific embodiments. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] in:

[0067] Figure 1 This is a schematic diagram of the workflow of a deep learning-based SOFC state monitoring and health diagnosis system provided in an exemplary embodiment of the present invention.

[0068] Figure 2 This is a schematic diagram of a CNN-LSTM model structure provided in an exemplary embodiment of the present invention;

[0069] Figure 3 This is a schematic diagram of the training accuracy curve provided by an exemplary embodiment of the present invention;

[0070] Figure 4 This is a comparison curve of HIS changes over time provided by an exemplary embodiment of the present invention. Detailed Implementation

[0071] 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. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0072] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0073] To facilitate a clearer and more accurate understanding of the technical solutions of this invention by those skilled in the art, the following will first provide a more detailed explanation of the technical problems existing in the prior art in conjunction with examples.

[0074] Because SOFCs operate at high temperatures (e.g., 600°C to 1000°C) and in complex chemical environments, they are susceptible to factors such as material degradation, thermal stress, and uneven fuel distribution, which can lead to performance degradation or even system failure.

[0075] In existing technologies, SOFC condition monitoring mainly relies on physical models (e.g., thermodynamic and electrochemical analysis models) and simple threshold judgments, such as providing early warnings by monitoring voltage drops or temperature anomalies. These condition monitoring methods were effective in early SOFC applications; however, as the complexity of SOFC systems increases and application scenarios expand, their limitations have become increasingly apparent.

[0076] In recent years, with the development of artificial intelligence and data processing technology, deep learning methods have begun to be introduced into the field of SOFC condition monitoring and fault diagnosis. For example, Mohamadali et al. (2024) proposed a deep learning method based on the HY-CNN-NARX paradigm for transient modeling of SOFC, demonstrating the potential of neural networks in dynamic performance analysis [1]. Zhang et al. (2021) developed a deep learning-based intelligent diagnostic method to achieve simultaneous fault identification of SOFC systems, promoting the intelligent process of SOFC health management [2]. In terms of patent technology, Chinese patent document CN112373352B proposes a fault diagnosis and fault-tolerant control method for fuel cell systems [5], while Chinese patent document CN113594510B designs a SOFC fault diagnosis scheme based on a cloud platform, introducing cloud computing technology [6]. The review by Ren et al. (2022) comprehensively reviews the development of SOFC control technology, providing background reference for monitoring and diagnosis [3].

[0077] Despite the progress made in the aforementioned research and technologies, SOFC condition monitoring and health diagnosis still face many problems and challenges, mainly including the following aspects:

[0078] First, the complexity of the data and the difficulty of feature extraction.

[0079] The data generated by SOFC operation is high-dimensional, time-varying, and strongly coupled, making it difficult for traditional physical models and shallow machine learning methods to effectively extract its spatiotemporal features. For example, traditional threshold-based methods cannot capture weak abnormal signals in dynamic changes, while the HY-CNN-NARX paradigm in reference [1] is good at modeling but does not focus on diagnosis.

[0080] Second, the diagnostic accuracy and robustness are insufficient.

[0081] Existing methods are prone to false alarms or missed alarms when multiple faults coexist or when operating conditions are complex. Although the deep learning methods in references [2] and [4] have improved diagnostic capabilities, their robustness in dealing with noise interference and high-dimensional data still needs to be improved. The traditional diagnostic method in Chinese patent document CN112373352B [5] lacks adaptability to complex fault modes.

[0082] Third, it lacks real-time performance.

[0083] Faults in SOFC operation often develop rapidly, and traditional monitoring systems have long response times, making it difficult to achieve early warning. For example, the cloud platform solution in Chinese patent document CN113594510B is affected by network latency, limiting its real-time performance, while the control technology mentioned in reference [3] has not solved the real-time problem of diagnosis.

[0084] Fourth, there is a lack of full lifecycle management.

[0085] Current technologies are mostly focused on single aspects (such as fault detection), lacking a closed-loop system from data acquisition to diagnosis and then to optimization feedback, which makes it impossible to achieve intelligent and continuous health management of SOFCs.

[0086] Therefore, designing a novel SOFC condition monitoring and health diagnosis solution to at least partially address the aforementioned issues has become an urgent problem to be solved.

[0087] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0088] According to a first aspect of the present invention, this embodiment provides a deep learning-based method for SOFC state monitoring and health diagnosis, comprising the following steps:

[0089] Step S1: Collect SOFC operating data, including temperature, total stack voltage, operating current, hydrogen flow rate, oxygen flow rate, and SOFC internal pressure;

[0090] Step S2: Preprocess the collected operating data and extract features from the preprocessed operating data. The features include voltage attenuation rate, voltage mean, voltage standard deviation, and voltage peak value.

[0091] Step S3: Input the preprocessed running data into the deep learning diagnostic model. The deep learning diagnostic model adopts a hybrid architecture of convolutional neural network and long short-term memory network, and combines an attention mechanism to output an adaptive health status index. ;

[0092] in, The calculation process specifically includes:

[0093] Step S3-1: Calculate the basic adaptive health status index :

[0094]

[0095] Step S3-2: Calculate the rate of change:

[0096]

[0097] Step S3-3: Adjust the weights:

[0098]

[0099] Step S3-4: Calculation :

[0100]

[0101] In the formula, First, set the weights. Set a weight for the second one. Voltage attenuation rate, For SOFC operating temperature, For the optimal operating temperature of SOFC, The maximum allowable operating temperature for SOFC. The adaptive health status index of the previous moment is used to calculate the rate of change and reflect the temporal correlation of the state. Adjust the weight first. The second adjustment weight;

[0102] Step S4: According to Real-time monitoring of its rate of change, if Below the preset threshold or If the rate of change exceeds a preset threshold, an early warning will be issued.

[0103] Through the above technical solution, firstly, this invention is the first to combine Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM) to design a novel deep learning model for multidimensional temporal series data of SOFC. CNN is used to extract spatial features, while LSTM is used to capture temporal dependencies. Compared to the existing HY-CNN-NARX paradigm and intelligent diagnostic methods, the hybrid model of this invention is superior in spatiotemporal feature analysis and can achieve higher accuracy in health status assessment.

[0104] Second, the adaptive health status index proposed in this invention The algorithm, by dynamically adjusting parameter weights, can quantify the health status of SOFC in real time. Compared with existing multi-label diagnostic methods, this algorithm can self-optimize according to changes in operating conditions, effectively improving the flexibility and accuracy of diagnosis.

[0105] Third, the method of this invention integrates data acquisition, preprocessing, diagnosis, and early warning steps, enabling the effective and accurate construction of a closed-loop management system for the entire SOFC lifecycle. Compared with traditional diagnostic methods and cloud platform solutions in existing related technologies, the closed-loop design of this invention can effectively reduce manual intervention and significantly improve the system's intelligence level and operational continuity.

[0106] Fourth, the method of this invention, through data augmentation techniques and attention mechanisms, can effectively improve the robustness of the model in complex environments. Simultaneously, by combining algorithm optimization and hardware acceleration, this invention can achieve rapid response (e.g., within 0.5 seconds) and a fault prediction window of several hours, far exceeding the real-time performance of existing related technologies.

[0107] In summary, this invention introduces a deep learning diagnostic model and an adaptive health status index. The closed-loop system covering the entire lifecycle can effectively solve key problems such as data complexity, insufficient accuracy, and poor real-time performance in SOFC condition monitoring and health diagnosis, providing an innovative technical solution for the efficient operation and long-life management of SOFC.

[0108] The method of the present invention will be described below with reference to an exemplary embodiment.

[0109] I. Experimental fuel cell stack:

[0110] The implementation of this invention relies on sophisticated experimental equipment and controlled environmental conditions to ensure the accuracy of data acquisition and the stability of the system.

[0111] The following are the detailed equipment and environment configurations:

[0112] The SOFC stack (using a 1kW planar solid oxide fuel cell (SOFC) stack, consisting of 20 individual cells connected in series). Among them,

[0113] The specifications for a single battery are as follows:

[0114] Active area: 100cm²;

[0115] Anode material: Nickel-zirconium oxide (Ni-YSZ), approximately 500 μm thick;

[0116] Cathode material: Lanthanum strontium manganese oxide (LSM), approximately 50 μm thick;

[0117] Electrolyte: Zirconia (YSZ), approximately 150 μm thick;

[0118] Collected runtime data:

[0119] Temperature: 800℃, maintained by a built-in ceramic heater, with a fluctuation range of ±5℃;

[0120] Fuel: Pure hydrogen (99.99% purity), flow rate range 0-10 L / min;

[0121] Oxidizing agent: Air (containing 21% oxygen), flow rate range 0-5L / min;

[0122] Initial voltage: 1.2V (approximately 0.06V for a single cell);

[0123] Current range: 0-50A;

[0124] Fuel supply:

[0125] Hydrogen is supplied through a high-pressure steel cylinder, and the pressure is adjusted to 0.1 MPa via a pressure reducing valve;

[0126] Air is supplied by a compressor, filtered, and then enters the fuel cell stack.

[0127] II. Implementation Steps

[0128] The implementation of this invention can be divided into four steps, from data acquisition to operation optimization. The specific operation, technical parameters and implementation details of each step are described below.

[0129] Step 1: Data Collection

[0130] The operating time is 30 consecutive days (i.e., 720 hours); the collected data includes temperature (anode inlet, cathode outlet and stack center, in °C), voltage (total stack voltage, in V), current (operating current, in A), flow rate (hydrogen flow rate and air flow rate, in L / min), pressure (internal system pressure, in kPa), and sampling frequency (1 Hz, i.e., once per second, for a total of approximately 25.92 million data points).

[0131] Operating conditions include: normal operation (voltage 1.2V, current 20A, temperature 800℃), load switching (current increases from 20A to 30A for 12 hours), and aging stage (voltage gradually decreases to 1.15V on day 25).

[0132] The purpose of this step is to obtain multidimensional data on the entire lifecycle of SOFC, providing a foundation for subsequent analysis.

[0133] Step 2: Data Preprocessing

[0134] Denoising was performed (using wavelet transform with a decomposition layer of 5); the calculation formula is as follows: In the formula, The standard deviation of noise. The standard deviation coefficient of the noise. , For signal length, (Data points per minute); to remove sensor noise (e.g., voltage jitter ±0.01V) and preserve operating trends.

[0135] Normalization (using linear normalization); the calculation formula is as follows: In the formula, For the normalized data, The original data, The minimum value in the original data. The maximum value in the original data is used; the parameter ranges are: temperature (600-1000℃), voltage (0V-1.2V), current (0A-50A), hydrogen flow rate (0L / min-10L / min), air flow rate (0L / min-5L / min), and pressure (0kPa-200kPa); to standardize the data to 0-1 and eliminate dimensional differences.

[0136] Feature extraction (extracting voltage decay rate, mean voltage, standard deviation of voltage, and peak voltage respectively):

[0137] The formula for calculating the voltage attenuation rate is as follows: In the formula, Voltage attenuation rate, The initial voltage, The current voltage is used to quantify performance degradation; for example, ; ; .

[0138] The formula for calculating the average voltage is: In the formula, For the sample size, The voltage of each individual cell in the fuel cell stack is used to represent the overall level; for example, The voltage data within a minute [1.18, 1.19, ..., 1.17] yielded a calculated average voltage of 1.18V.

[0139] The formula for calculating the standard deviation of voltage is: This is used to measure the magnitude of voltage fluctuations; for example, the standard deviation is calculated to be 0.008V.

[0140] For the peak voltage, the calculation formula is: This is used to capture abnormal voltage values; for example, the peak value within a minute is 1.19V.

[0141] Step 3: Model Training (For accuracy changes during training, please refer to...) Figure 3 ),exist Figure 3 The graph shows the changes in training and validation accuracy. The curve illustrates the performance of the CNN-LSTM model during training, with the horizontal axis representing the number of training epochs (1 to 100) and the vertical axis representing accuracy (0 to 1). The solid blue line represents training accuracy, and the dashed orange line represents validation accuracy. Both curves gradually stabilize with increasing epochs, indicating strong learning ability and no significant overfitting. Grid lines and legends enhance the readability of the curve, demonstrating the model's convergence and high performance (e.g., accuracy close to 1), providing intuitive validation of the system's prediction accuracy.

[0142] 1) Regarding the model structure, where,

[0143] CNN (Convolutional Neural Network):

[0144] First layer: 32 3×3 convolutional kernels, stride 1, padding "same", activation function ReLU;

[0145] Second layer: 64 3×3 convolutional kernels, stride 1, padding "same", activation function ReLU;

[0146] Pooling layer: 2×2 max pooling, reducing features to 30×64;

[0147] Input: 60×6 (60 seconds × 6 parameters);

[0148] Output: Time series feature vector;

[0149] LSTM (Long Short-Term Memory Network):

[0150] First layer: 128 units, returns a sequence, Dropout rate 0.2;

[0151] Second layer: 64 units, returns a single vector, Dropout rate 0.2;

[0152] Input: CNN output (30×64);

[0153] Output: 64-dimensional temporal features;

[0154] Attention mechanism:

[0155] The calculation formula is: , to be used to measure feature importance; where, To calculate the score, Output for LSTM (64-dimensional). These are trainable weight vectors;

[0156] Output: Weighted feature vector;

[0157] 2) The training parameters are as follows: input dimension 60×6 (60 seconds × 6 parameters), batch size 32, learning rate 0.001, training epochs 100, and loss function is mean squared error (MSE). In the formula, To represent the true experimental values, For the predicted value, the optimizer is ),in, The exponential decay rate is estimated by the first moment. The exponential decay rate is estimated by the second moment. A very small number is used for numerical stability to prevent division by zero in calculations; the data is divided into an 80% training set (576 hours) and a 20% validation set (144 hours).

[0158] 3) The training results are:

[0159] Validation set performance: HSI prediction error <5%, correlation coefficient 0.95 (HSI vs. actual decay rate), MSE <0.01.

[0160] Step 4: Running and Optimizing

[0161] Among them, the monitoring mechanism is as follows:

[0162] calculate:

[0163] The calculation formula is: Among them, the parameters are: , Initial weights: , ;

[0164] Dynamic adjustment:

[0165] Calculate the rate of change: ;

[0166] like , ;

[0167] Warning conditions:

[0168] Rate of change: Threshold: rate > 0.05;

[0169] Output: "Advanced Warning";

[0170] Threshold determination: ;

[0171] Output: "Audio and visual alarm";

[0172] Optimization strategy:

[0173] When HSI < 0.9, reduce the current;

[0174] When the temperature is >900℃, increase the cooling air by 10%;

[0175] Optimization effect:

[0176] The HSI rose to 0.88 after the temperature anomaly.

[0177] To prevent further damage after voltage decay.

[0178] III. Experimental Results and Invention Effects

[0179] Experimental results

[0180] Case 1: Abnormal temperature;

[0181] Time: Day 15, 360 hours;

[0182] Scenario: Temperature rises from 800℃ to 920℃, HSI drops from 0.95 to 0.85;

[0183] System response:

[0184] Detected 2 hours in advance (358th hour) ;

[0185] An "advanced warning" is issued, suggesting an increase in cooling airflow of 10%.

[0186] Result: The temperature dropped to 810℃, and the HSI rose back to 0.88, avoiding damage from overheating.

[0187] Case 2: Voltage decay;

[0188] Time: Day 25, 600th hour;

[0189] Scenario: Voltage drops from 1.2V to 1.14V (ΔV=8% Delta V = 8%), HSI drops to 0.72;

[0190] System response:

[0191] Trigger the "audio and visual alarm";

[0192] It is recommended to reduce the current to 80%;

[0193] Result: During shutdown and maintenance, electrode aging was discovered, preventing further damage to the fuel cell stack.

[0194] Based on the above exemplary embodiments, the advantages of the present invention include:

[0195] First, it enables high-precision monitoring. Specifically, the HSI prediction error is <5%. The method of this invention, through a CNN-LSTM model and attention mechanism, can effectively and accurately predict the health status of SOFCs, with a correlation coefficient of 0.95 with the actual decay rate. Compared with traditional thresholding methods (error 15%-20%) and single CNNs (error 8%-10%), the method of this invention has better prediction accuracy and can effectively ensure the detection of early anomalies.

[0196] Second, it can meet real-time requirements (see reference). Figure 4The system can respond in advance to abnormal temperatures within 50 hours. Specifically, the average response time of the method of this invention is 0.42 seconds (<0.5 seconds), and rapid response can be achieved through hardware acceleration and algorithm optimization. Simultaneously, it can provide early warnings 2-4 hours in advance (e.g., detecting a 920℃ abnormality 2 hours in advance). Compared to the several-minute delay of traditional methods, the method of this invention has strong real-time performance and can provide sufficient maintenance time.

[0197] Among them, Figure 4 The curve compares the HSI changes of the present invention with those of traditional methods. The curve shows the change in the SOFC Health Status Index (HSI) over 100 hours of operation, with the horizontal axis representing operation time (hours) and the vertical axis representing HSI (0 to 1). The solid line represents the prediction result of the present invention (CNN-LSTM + Adaptive HSI), the dashed line represents the traditional thresholding method, and the dotted line represents a single CNN model. The curve simulates scenarios where the voltage linearly decreases from 1.2V (0.001V per hour) and the temperature rises from 800℃ to 850℃ over 50 hours. The present invention's HSI decreases smoothly, responding to the temperature anomaly (850℃) 50 hours in advance. The traditional thresholding method drops sharply to 0.7 when the voltage is <1.0V or the temperature is >840℃ at 60 hours, showing significant lag. The single CNN model exhibits large fluctuations due to noise (error of approximately 10%). Grid lines and legends enhance readability, demonstrating the advantages of the present invention in accuracy (smooth prediction) and real-time performance (early response), providing intuitive evidence of the present invention's technical advantages.

[0198] Third, the method of the present invention has better robustness. Specifically, the noise resistance of the method of the present invention can be significantly improved (e.g., 10%), while the data augmentation and denoising techniques can ensure that the HSI prediction error is <5% under complex environments (e.g., 80% humidity). Thus, the method of the present invention can adapt to noise and operating condition fluctuations in industrial scenarios and reduce misjudgments.

[0199] Fourth, the method of this invention can effectively improve the lifespan and efficiency of SOFCs. Specifically, by real-time optimization (e.g., reducing current, increasing cooling) to slow down the aging process, the lifespan of the SOFC can be effectively extended, for example, by 15% (from 2400 hours to 2760 hours). Simultaneously, the method of this invention can optimize operating parameters to improve energy conversion efficiency, and the efficiency of the SOFC can also be effectively improved, for example, by 5%-10% (from 85% to 90%). This effectively extends the service life of the equipment and reduces maintenance costs.

[0200] Fifth, the method of this invention enables closed-loop management. Specifically, it enables full lifecycle management from data collection to optimization recommendations. Compared to traditional segmented monitoring, it offers a higher degree of automation and reduces manual intervention.

[0201] Sixth, in specific fields, the method of this invention can achieve high application value. For example, in the field of industrial power generation, it can effectively improve the reliability of SOFCs and reduce downtime losses. In the aerospace field, it can achieve real-time early warning to ensure the safety of critical missions. Simultaneously, in many application areas, it can extend lifespan, improve efficiency, and reduce total cost of ownership.

[0202] In one embodiment of the present invention, the deep learning-based SOFC state monitoring and health diagnosis method of the present invention may further include:

[0203] Step S5: In response to the warning results, display the health status of SOFC through a visual interface and provide optimization suggestions.

[0204] In this way, through this step S5, the health status of SOFC can be intuitively displayed on the visualization interface based on the warning results, and targeted optimization suggestions can be provided.

[0205] In one embodiment of the present invention, step S5 specifically includes: in response to the warning result, displaying the health status of SOFC through a visual interface;

[0206] Output the corresponding optimization suggestions based on the following formula:

[0207]

[0208] In the formula, To optimize the suggestions, This refers to the actual operating temperature of the SOFC.

[0209] In this way, a "monitoring-diagnosis-optimization" closed loop is achieved through a dynamic feedback mechanism. Specifically, when HSI < 0.9, current regulation (load derating protection) is automatically triggered, and when the temperature exceeds 900℃, the cooling system is automatically adjusted (thermal management), forming a real-time parameter optimization closed loop.

[0210] In one embodiment of the present invention, step S2 may specifically include:

[0211] Step S2-1: Denoise the collected running data based on wavelet transform;

[0212] Step S2-2: Normalize the noise-reduced running data;

[0213] Step S2-3: Extract features from the normalized running data.

[0214] In this implementation, wavelet transform is used for multi-scale decomposition, effectively distinguishing high-frequency noise from real-world operating condition fluctuation signals, thereby improving the signal-to-noise ratio. Simultaneously, normalization processing establishes a unified dimensional space, eliminating the interference of magnitude differences on model training.

[0215] In one embodiment of the present invention, the deep learning diagnostic model of the present invention includes a CNN part and an LSTM part; the CNN part includes two convolutional layers, each convolutional layer is followed by a ReLU activation function, and the dimensionality is reduced by a max pooling layer after the convolution operation; the LSTM part includes two LSTM layers, each LSTM layer is followed by a Dropout layer to prevent overfitting.

[0216] The structure of the deep learning diagnostic model can be found in [reference]. Figure 2 As shown, in this embodiment, the formula for the convolution operation of the deep learning diagnostic model is:

[0217]

[0218] In the formula, This is the output value at coordinates (i,j) in the feature map output by the convolutional layer. These are the horizontal and vertical indices of the convolution kernel, used to traverse the elements of the kernel. The values ​​range from 0 to 2, indicating a kernel size of 3×3. The input feature value is the feature value at coordinates (i+m, j+n) in the feature map. Let be the weight value at coordinate (mn) in the convolution kernel. This is used as a bias to extract spatial features.

[0219] In LSTM updates, the formula for calculating the forget gate is: Its function is to control history retention. The calculation formula for the input gate is: Its function is to control the addition of new information. The formula for calculating candidate states is: Its function is to generate updated content. The formula for calculating the output gate is: Its function is to control the current output. The formula for calculating the updated cell state is: Its function is to record long-term memories. The formula for generating hidden states is: Its function is to output the analysis results.

[0220] In the formula, This is the output of the forget gate. It is a value between 0 and 1. It is the Sigmoid function, which compresses function values ​​to the range of 0 to 1. , , , These are the forget gate and the weight matrix. This indicates that the previous state will be hidden. Input at the current time Piecing them together, It is a bias term. It is the cell state at time t. It is a candidate state at time t. It is the output of the output gate.

[0221] This invention is the first to combine Convolutional Neural Networks (CNNs) with Long Short-Term Memory Networks (LSTMs), designing a two-layer CNN (32 3×3 convolutional kernels in the first layer, 64 in the second layer) and a two-layer LSTM (128 units in the first layer, 64 units in the second layer), specifically optimized for the multidimensional characteristics (including temperature, voltage, current, flow rate, etc.) and time-varying characteristics (performance degradation over time) of SOFC operating data. The CNN extracts spatial features through convolutional operations, while the LSTM analyzes the time series through a gating mechanism, outputting a Health Status Index (HSI).

[0222] Compared to single CNNs or traditional machine learning methods (such as SVM), hybrid models significantly improve the ability to capture the spatiotemporal correlations of SOFC data. In experiments, the HSI prediction error was reduced to below 5%, with a correlation coefficient of 0.95, while the error of a single CNN was approximately 8%-10%, and SVM could only handle static data with an accuracy of less than 70%. For example, when voltage decreases slowly while temperature fluctuates, hybrid models can comprehensively analyze the long-term effects of both, while traditional methods may ignore the time effect.

[0223] This innovation improves the accuracy and depth of SOFC condition monitoring, addressing the shortcomings of traditional methods in identifying complex operating modes. For example, in distributed generation, the system can accurately identify early signs of performance degradation, preventing large-scale failures. This technological breakthrough is unprecedented in the field of SOFC monitoring, filling a gap in the application of deep learning in fuel cell dynamic analysis.

[0224] In one embodiment of the present invention, the attention mechanism of the present invention is implemented in the following way:

[0225] Step S3-2-1: Calculate the feature vector score output by the Long Short-Term Memory network;

[0226] Step S3-2-2: Normalize the feature vector scores;

[0227] Step S3-2-3: Weight key features to highlight their importance. The contribution of the assessment.

[0228] In this implementation, a dynamic feature weighting system based on an attention mechanism is constructed, which can significantly improve the ability to capture key fault features in SOFC condition monitoring. Specifically, firstly, a learnable feature score calculation module is used to automatically generate attention weights from the temporal feature vector output by LSTM, which can effectively improve the accuracy of sensitive feature identification compared to the traditional fixed weight allocation method. Secondly, by normalizing the feature vector scores, irrelevant noise interference can be effectively suppressed. Thirdly, key features can be weighted, thereby achieving dynamic weight adjustment.

[0229] Key characteristics may include battery voltage, flow rate, temperature, and pressure.

[0230] In one embodiment of the present invention, step S4 may specifically include:

[0231] Calculate according to the following formula Decline rate:

[0232]

[0233] Where, The adaptive health status index at time tk. This is the time interval (i.e., the step size).

[0234] when When the value is greater than the first preset threshold, the first warning signal is output;

[0235] when When the value is less than the second preset threshold, a second warning signal is output.

[0236] Thus, in this embodiment, by The formula dynamically tracks performance degradation trends, which, compared to fixed threshold monitoring methods, can improve the success rate of detecting early-stage performance degradation in SOFC.

[0237] In one embodiment of the present invention, the present invention The value is 5, that is, The first preset threshold is set to 0.05; and / or the second preset threshold is set to 0.7.

[0238] According to a second aspect of the invention, such as Figure 1As shown, a deep learning-based SOFC state monitoring and health diagnosis system is also provided, used to execute the deep learning-based SOFC state monitoring and health diagnosis method of any of the technical solutions in the first aspect of the present invention. The deep learning-based SOFC state monitoring and health diagnosis system includes a data acquisition module, a preprocessing module, a deep learning diagnostic model, and a state monitoring and early warning module. The data acquisition module is used to collect SOFC operating data. The preprocessing module is used to preprocess the collected operating data and extract features from the preprocessed operating data. The deep learning diagnostic model adopts a hybrid architecture of convolutional neural network and long short-term memory network, combined with an attention mechanism. The deep learning diagnostic model is used to analyze the data and output an adaptive health status index. The status monitoring and early warning module is used to... It monitors and issues early warnings in real time about its rate of change.

[0239] In one embodiment of the present invention, the deep learning-based SOFC state monitoring and health diagnosis system may further include a visualization and feedback module. The visualization and feedback module is used to respond to early warning results by displaying the health status of the SOFC through a visual interface and providing optimization suggestions.

[0240] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning-based method for SOFC state monitoring and health diagnosis, characterized in that, Includes the following steps: Step S1: Collect SOFC operating data, including temperature, total stack voltage, operating current, hydrogen flow rate, oxygen flow rate, and SOFC internal pressure; Step S2: Preprocess the collected operating data and extract features from the preprocessed operating data. The features include voltage attenuation rate, voltage mean, voltage standard deviation, and voltage peak value. Step S3: Input the preprocessed running data into the deep learning diagnostic model. The deep learning diagnostic model adopts a hybrid architecture of convolutional neural network and long short-term memory network, and combines an attention mechanism to output an adaptive health status index. ; in, The calculation process specifically includes: Step S3-1: Calculate the basic adaptive health status index : Step S3-2: Calculate the rate of change: Step S3-3: Adjust the weights: Step S3-4: Calculation : Where, First, set the weights. Set a weight for the second one. Voltage attenuation rate, For SOFC operating temperature, For the optimal operating temperature of SOFC, The maximum allowable operating temperature for SOFC. The adaptive health status index of the previous moment is used to calculate the rate of change and reflect the temporal correlation of the state. Adjust the weight first. The second adjustment weight; Step S4: According to Real-time monitoring of its rate of change, if Below the preset threshold or If the rate of change exceeds a preset threshold, an early warning will be issued.

2. The deep learning-based SOFC state monitoring and health diagnosis method according to claim 1, characterized in that, The deep learning-based SOFC state monitoring and health diagnosis method also includes: Step S5: In response to the warning results, display the health status of SOFC through a visual interface and provide optimization suggestions.

3. The deep learning-based SOFC state monitoring and health diagnosis method according to claim 2, characterized in that, Step S5 specifically includes: In response to the warning results, the health status of SOFC is displayed through a visual interface; Output corresponding optimization suggestions: In the formula, To optimize the suggestions, This refers to the actual operating temperature of the SOFC.

4. The deep learning-based SOFC state monitoring and health diagnosis method according to claim 1, characterized in that, Step S2 specifically includes: Step S2-1: Denoise the collected running data based on wavelet transform; Step S2-2: Normalize the noise-reduced running data; Step S2-3: Extract features from the normalized running data.

5. The deep learning-based SOFC state monitoring and health diagnosis method according to claim 1, characterized in that, The deep learning diagnostic model includes a CNN part and an LSTM part; The CNN part includes two convolutional layers, each followed by a ReLU activation function, and then dimensionality reduction is performed through a max pooling layer after the convolution operation. The LSTM part includes two LSTM layers, each followed by a Dropout layer to prevent overfitting.

6. The deep learning-based SOFC state monitoring and health diagnosis method according to claim 1, characterized in that, The attention mechanism is implemented in the following ways: Step S3-2-1: Calculate the feature vector score output by the Long Short-Term Memory network; Step S3-2-2: Normalize the feature vector scores; Step S3-2-3: Weight key features to highlight their importance. The contribution of the assessment.

7. The deep learning-based SOFC state monitoring and health diagnosis method according to claim 1, characterized in that, Step S4 specifically includes: Calculate according to the following formula Decline rate: In the formula, for A real-time adaptive health status index. For time intervals; when When the value is greater than the first preset threshold, the first warning signal is output; when When the value is less than the second preset threshold, a second warning signal is output.

8. The deep learning-based SOFC state monitoring and health diagnosis method according to claim 7, characterized in that, The The value is 5, and the first preset threshold is set to 0.05; and / or, The second preset threshold is set to 0.

7.

9. A deep learning-based SOFC state monitoring and health diagnosis system, characterized in that, For executing the deep learning-based SOFC state monitoring and health diagnosis method as described in any one of claims 1-8, the deep learning-based SOFC state monitoring and health diagnosis system comprises: The data acquisition module is used to collect SOFC's runtime data; The preprocessing module is used to preprocess the collected running data and extract features from the preprocessed running data. The deep learning diagnostic model employs a hybrid architecture of convolutional neural networks and long short-term memory networks, combined with an attention mechanism. This model is used to analyze data and output an adaptive health status index. ; The status monitoring and early warning module is used to... It monitors and issues early warnings in real time about its rate of change.

10. The SOFC state monitoring and health diagnosis system based on deep learning according to claim 9, characterized in that, The deep learning-based SOFC state monitoring and health diagnosis system also includes: The visualization and feedback module is used to respond to the warning results, display the health status of SOFC through a visual interface, and provide optimization suggestions.

Images