TWI395965B - Fuel cell faulty predicting system and its establishing method - Google Patents

Description

Fuel cell fault prediction system and its establishment method

The present disclosure relates to a monitoring device, and more particularly to a fuel cell monitoring device.

Fuel cells are environmentally friendly because they can be recharged. However, when a fuel cell is to be used in a large number of applications, it is necessary to ensure its reliability first; in important industrial applications, we can only build a redundant structure for the fuel cell to realize the concept of uninterrupted power; The cost is a set of battery arrays that use inefficiencies. On the other hand, a variety of fuel cell monitoring mechanisms have also been proposed in order to detect fuel cell failures in a timely manner, and to avoid further losses caused by power outages.

Accordingly, it is a technical aspect of the present disclosure to provide a fuel cell failure prediction system that can detect or repair a component that may be malfunctioning in the future when the fuel cell has not failed.

According to an embodiment of the present invention, a fuel cell fault prediction system is provided, including a plurality of detectors, a digital signal processing unit, a gray prediction operation unit, an extension type neural network classifier, and a display. The detector is used to monitor multiple status signals of a fuel cell system. The digital signal processing unit is configured to generate at least one set of characteristic signals according to the plurality of status signals. The gray prediction operation unit is configured to generate a set of gray prediction signals according to the set of characteristic signals. The extensional neurological operation unit is trained by using multiple fault models to generate a detection data according to the characteristic signal, and generate a prediction data according to the gray prediction signal. The display is used to display the detected data and forecast data.

It should be noted that the technical aspect of the present invention further includes wirelessly transmitting the feature signal to the gray prediction operation unit and the extension type neural network classifier by using a wireless signal transceiver.

Another aspect of the present disclosure is to provide a fuel cell fault prediction system establishment method for obtaining various fault data without actually destroying a plurality of sets of fuel cells, thereby establishing the aforementioned fuel cell fault prediction system.

According to an embodiment of the present technology, a method for establishing a fuel cell fault prediction system is provided, including the following steps: First, a plurality of detectors are provided to monitor a plurality of status signals of a fuel cell system. Then, a fuel cell model is established by using the state signal, and then the fuel cell model is used to generate a plurality of sets of fault characteristic signals. Next, a plurality of matter-element models are built according to the fault feature signals, and the matter-element models are used to train an extension-like neural network system. On the other hand, these fault signature signals are also used to create a gray prediction model. Finally, the gray prediction model is set in front of the extension type neural network system, whereby the gray prediction model can generate a prediction feature signal to the extension type neural network system according to one of the received measured characteristic signals.

It should be noted that the present technical aspect, in other embodiments, proposes to establish a fuel cell model including establishing an electrochemical submodel and a thermodynamic submodel. Electrical sub-model to describe the system activated anode and the cathode within the fuel cell arising from a voltage drop V _act, describe the ohmic drop within the fuel cell voltage V _ohmic, and the losses described by the concentration of mass transfer due to diffusion limitations within the fuel cell V _Con . On the other hand, the thermodynamic submodel is used to describe the reversible voltage E _{thermo of the} fuel cell output thermodynamics. In addition, the matter element model includes a fan failure matter element model, a heat dissipation system fault matter element model, a fuel consumption oversized matter element model, a hydrogen pressure shortage matter element model, a vent hole blockage element model, and an overload matter element. model. Moreover, each matter element model includes a hydrogen pressure fault characteristic value, a fuel cell operating temperature characteristic value, a fuel cell voltage characteristic value, a fuel cell current characteristic value, an input air flow characteristic value, and an exhaust port relative Humidity characteristic value.

In this way, the fuel cell fault prediction system of the foregoing embodiments can detect the current signal and determine the fault category; and, based on the current signal, generate a prediction signal to determine the fault category that may occur in the future.

Please refer to FIG. 1. FIG. 1 is a functional block diagram of a fuel cell fault prediction system according to an embodiment of the present disclosure. In the first figure, the fuel cell fault prediction system includes a plurality of detectors 110, a digital signal processing unit 120, a gray prediction operation unit 130, an extension type neural network classifier 140, and a display 150. The detector 110 is installed in various parts of a fuel cell system 100 to monitor a plurality of status signals of the fuel cell system 100. The digital signal processing unit 120 is configured to generate at least one set of characteristic signals according to the plurality of status signals. The gray prediction operation unit 130 is configured to generate a set of gray prediction signals according to the set of characteristic signals. The extension-type neural network classifier 140 is trained by using multiple fault models to generate a detection data according to the feature signal, and generate a prediction data according to the gray prediction signal. The display 150 is used to display the detected data and the predicted data. The construction method and operation principle of the above components are as follows: Please refer to FIG. 2, and FIG. 2 is a detailed structural diagram of the fuel cell system 100 of FIG. 1. For convenience of explanation, the digital signal processing unit 120 is further illustrated. Other components. For the fuel cell system 100, common causes of failure can be divided into six categories, namely, fan failure, heat dissipation system failure, excessive fuel consumption, insufficient hydrogen pressure, clogging and overload of the vent. When any of the above six types of faults occur, numerical changes occur in the following six states: hydrogen pressure, operating temperature, battery voltage, battery current, input air flow, and exhaust port relative humidity. Therefore, the plurality of detectors 110 are installed at positions where at least the above listed state can be detected to monitor changes in the fuel cell system 100. The detector 110 can include a pressure detector 111, a flow detector 112, and a battery detector 113. The battery detector 113 is used to detect the voltage, current, temperature, and drain of the fuel cell. Humidity, and the plurality of status signals detected by the detectors 110 are uniformly digitized by the digital signal processing unit 120 for subsequent processing. Further, in other embodiments, the digital signal processing unit 120 can transmit the digitized status signal to a group of wireless signal transceivers 160, such as ZigBee, and then to the gray prediction operation unit 130 by the wireless signal transceiver 160. The extension-type neural network classifier 140 and the display 150 are usually a computer 170.

Please refer to FIG. 3A, FIG. 3B, FIG. 3C and FIG. 3D together. FIG. 3A is a schematic structural diagram of a neural network such as the extension type neural network classifier 140 of FIG. 1 , and FIG. 3B is a schematic diagram. The extension distance of Figure 3A shows FIG. 3C is a schematic diagram of the extension distance before the weight adjustment of FIG. 3B, and FIG. 3D is a schematic diagram of the extension distance after the weight adjustment of FIG. 3B. Specifically, the extension-like neural network has the adaptability of accepting different kinds of variables as inputs, and the 3A diagram includes the input layer, the calculation layer, and the output layer. The process of this architecture first classifies the input data and constructs the matter element model and then enters the extension class neural network. The number of input layers is determined by the number of features of the matter element model, while the output layer is composed of data. The number of categories determines and stores the calculated extension distance. Finally, the minimum value of the extension distance value (ED value) belonging to the output layer of each category determines the category of the data.

The learning rules of extension-like neural networks can be divided into unsupervised learning and supervised learning. Non-supervised learning is based on the current possessed feature sample values, and learning to find out the regularity of data. Relevance, and when there is a data to be input for identification, it is to find the most similar one as the identification result. The extension-type neural network learning method used in the present embodiment uses supervised learning. The supervised learning is to adjust the weight through learning, and the adjustment weights and identification are adjusted by continuous learning and training, thereby reducing the extension. The difference between the output value of the neural network and the target output value, thereby improving the accuracy of the extension neural network identification. Therefore, there must be a learning sample X={X ₁ , X ₂ , X ₃ ,...X _Pm } before learning, and each sample contains the characteristics and categories of the data X _i ^m ={X _i1 ^m ,X _i2 ^m , X _i3 ^m ,...X _in ^m }, the learning samples are represented by the symbol P, P _M is the total number of samples, and m is the total number of features. The total error is set to P _N and the total error ratio is set to E _T , E _T = P _N / P _M . The calculation steps of the extension-like neuromonitoring learning are as follows:

Step 1: Use the extension material element model to establish the weight values of the input and output, and the expression of the matter element is as follows: In the above formula, m represents the total number of categories of data, R _k represents the matter element model, N _k is the name of the thing, c _j is the nth feature in the matter element model, and j=1, 2, 3,..., n, For the classical domain of the feature c _j , the classical domain range can be determined by the learning materials, as follows: Among them, x _ij ^k represents the input learning data of the extension type neural network.

Step 2: Calculate the weight center value of each feature, expressed as Z _k , as follows: Z _k ={ z _{k 1} , z _{k 2} , z _{k 3} ,... z _kn }

If the learning data only has the same set of data in the same feature, because the upper weight limit is equal to the lower weight limit, the following two formulas are used to adjust to avoid the extension distance and thus generate an infinite value. : α is the classical domain range adjustment rate. When the α setting is larger, the classical domain range is also increased.

Step 3: Read the i-th training sample data and feature number k as follows:

Step 4: Start the calculation of the extension distance (ED) using x _i ^k as follows: The extension distance diagram derived from the above formula is shown in Fig. 3B. The extension distance can be used to represent the distance between the point x and the range < W ^L , W ^U 〉. It can be seen from Fig. 3B that when the classical domain range of the eigenvalue is larger, the range of the learning data is larger, and the extension distance is calculated at this time. The lower the sensitivity is, the opposite is true. If the classical domain of the eigenvalue is smaller, the more accurate the data sample is, the higher the sensitivity is.

Step 5: Find the minimum extension distance for all categories: The category of the minimum extension distance is judged as the category k*. If the k* category is the same as the data category k, ie k*=k, and jumps to step 7; if the data categories are not equal k*≠k, then continue Step 6 action.

Step 6: Adjust the weight values of the k category and the k* category.

(1) Update the upper and lower limits of the weights. The operation steps are as follows:

(2) Update the weight center value, the operation steps are as follows: Among them, η is the learning rate. The learning rate affects the convergence speed and the accuracy of convergence. If the learning rate is larger, it is easy to reach convergence, but the accuracy of convergence may also decrease. Conversely, if the learning rate is smaller, the convergence is more accurate, but the number of learning and time will increase. The diagram in the adjustment process is shown in Figures 3C and 3D. In Figure 3C, because ED _{k*_old} <ED _{k_old} , the category determined is not the type of data. _Thereafter , ED _{k*_new} >ED _{k_new} calculated as the learning material of the 3D map indicates that the category to which it belongs has been changed to the correct category by adjusting the weight.

Step 7: Repeat steps 3 through 7 until all learning materials are read and learned to complete the classification.

Step 8: Stop when all the classification procedures of the data have reached the convergence state or the total error rate reaches the target value, otherwise return to step 3 to continue.

After the above steps, the present embodiment learns and adjusts the weight value by using the supervised learning method of the extension type neural network, and the weight value and the recognition rate can be reduced by continuous learning and training. The difference between the output value of the extension neural network and the target output value can improve the accuracy of the extension neural network identification.

Please refer to Figure 4 and Figure 5 together. Figure 4 is the picture of Figure 3A. A flow chart of the learning steps of the extension neural network, and FIG. 5 is a flow chart of the operation steps of the extension type neural network of FIG. 3A. It can be clearly seen from Fig. 4 that the learning method of the extension-like nerve and how to end the learning, the learning can jump out of the circle as long as the diagnosis rate is diagnosed, and the fault of the extension-type neurological diagnosis fuel cell can be broken. Identification. It can be clearly seen from Figure 5 that the extensional nerves operate.

However, when establishing the above-described extension-type neural network classifier 140, it is necessary to obtain enough fault samples. If the destructive fault data collection is performed by the real fuel cell system 100, the cost and time are considerable. Therefore, another technical aspect of the present disclosure is to provide a fuel cell fault prediction system establishing method, which can obtain various fault data without actually destroying multiple sets of fuel cells, thereby establishing the aforementioned fuel cell fault prediction system.

According to an embodiment of the present technology, a method for establishing a fuel cell fault prediction system is provided, including the following steps: First, a plurality of detectors are provided to monitor a plurality of status signals of a fuel cell system. Then, a fuel cell model is established by using the state signal, and then the fuel cell model is used to generate a plurality of sets of fault characteristic signals. Next, a plurality of matter-element models are built based on the fault feature signals, and the matter-element models are used to train a class-extension neural network system. On the other hand, these fault signature signals are also used to create a gray prediction model. Finally, the gray prediction model is set in front of the extension neural network system, whereby the gray prediction model can generate a prediction feature signal to the extension type neural network system according to one of the received measured characteristic signals.

It should be noted that the present technical aspect, in other embodiments, proposes to establish a fuel cell model including establishing an electrochemical submodel and a thermodynamic submodel. Electrical sub-model to describe the system activated anode and the cathode within the fuel cell arising from a voltage drop V _act, describe the ohmic drop within the fuel cell voltage V _ohmic, and the losses described by the concentration of mass transfer due to diffusion limitations within the fuel cell V _Con . On the other hand, the thermodynamic submodel is used to describe the reversible voltage E _{thermo of the} fuel cell output thermodynamics. In addition, the matter element model includes a fan failure matter element model, a heat dissipation system fault matter element model, a fuel consumption oversized matter element model, a hydrogen pressure shortage matter element model, a vent hole blockage element model, and an overload matter element. model. Moreover, each matter element model includes a hydrogen pressure fault characteristic value, a fuel cell operating temperature characteristic value, a fuel cell voltage characteristic value, a fuel cell current characteristic value, an input air flow characteristic value, and an exhaust port relative Humidity characteristic value.

Specifically, the actual fuel cell output voltage is a thermodynamic output voltage minus the voltage loss caused by various consumptions: V = E _thermo + V _act + V _ohmic + V _con E _thermo represents the reversible thermodynamics of the fuel cell output Sex voltage: In the above formula versus It is the pressure of hydrogen and oxygen required for the fuel cell, and T is the operating temperature of the fuel cell.

V _act is the voltage drop due to the activation of the anode and cathode: In the above formula, ξ _i (i=1-4) is the characteristic coefficient of various fuel cells, and I _FC is the fuel cell current value. (Unit: atmospheric pressure / atm) is the oxygen concentration.

V _ohmic called ohmic voltage drop due to charge transfer resistance would result in a voltage loss of the fuel cell, which is to follow Ohm's law. If thinner the electrolyte membrane and the highly conductive material, the ohmic losses may be minimized fuel _{cells: V ohmic = - I FC (} R M + R C)

R _M = ρMl / A

In the above formula, R _C is the resistance of the contact electron flow, R _M is the impedance through the proton exchange membrane, l is the thickness, A is the area, and ψ is the selected material coefficient. For the derivation and demonstration of the equation, please refer to the book by Fyan O'Hayre, Suk-Won Cha, Whitney Colella, Fritz B. Prinz, FUEL CELL FUNDAMENTALS.

V _con represents the concentration loss caused by diffusion-limited mass transfer: V _con =- B ln[1-( J / J _MAX )] where B (unit: volt/V) is expressed as a constant for any type of fuel cell , J _MAX is the maximum current density, and J is the current density generated by the battery.

Therefore, after the fuel cell model is established, various fault states can be simulated, and then enough samples can be obtained to train the extension-type neural network classifier 140 without excessive cost and time for actual fault detection. .

Finally, the operation principle and establishment method of the gray prediction operation unit 130 are as follows: the gray system theory is mainly used to study the uncertainty of a small sample or a small amount of information, and the correlation degree using the gray theory can be used with valuable information. Solving the system problem of unknowable information, so the gray theory is often used in the state where information is difficult to collect and the environment is often fiercely changed. To complete the purpose of forecasting, we must first establish a model of prediction, that is, use historical data from the past, use statistics or build A modular approach that produces a set of mathematical models of predictions. Traditional prediction methods such as Time series method, statistical methods or recent artificial intelligence methods such as expert systems and neural networks require a large amount of historical data to obtain better prediction results, while in grey theory The gray prediction GM(1,1) only needs a small number of samples to guess the future numerical changes, and only needs 6 data to make predictions, because it can generate irregular index series after accumulating and generating index rules. Then, the differential equation is built up by accumulating the number of generations, so it can be modeled with a minimum of 6 sets of data. If the gray theory GM(1,1) is used as the main predictive model, the matrix of the original test data is X=(x(1) ), x(2),...,x(n)), according to the gray theory GM(1,1) model, the system change state can be described by using the first-order differential equation as follows: In order to improve the accuracy of the prediction, in the present embodiment, the original data is subjected to secondary accumulation generation (2-AGO), and then z is a value obtained by the second accumulation of the original data, which is defined as follows: Y = ( y (1) y (2) ... z ( n ))

Z =( z (1) z (2)... z ( n )) where: , In the above formula, x(t) is the data detected at the tth time. In the gray theory, the first-order differential equation of the above formula is called the whitening equation and can be used to describe the white system (or a very complete system). For gray systems or predictions Approximate as follows: as well as . In the above formula, the best solution of the two parameters a and b can be solved by the least square method. Where: Y =[ y (2) y (3)... y ( n )] ^T

Parameters â and After solving, accumulate the predicted value of the generated value for: Predicted value of measurement data Can be obtained by using the following formula: With the above formula, the gray prediction operation unit 130 of the present embodiment can predict the next feature data of the system by only 6 pieces of data, and then use the completed extension type neural network classifier 140 to predict the fuel cell system in advance. Whether the critical point of failure has been reached, and the type of failure that may occur is further predicted.

Finally, please refer to Figure 6, which is a flow chart of the working steps of the fuel cell fault prediction system of Figure 1. In the sixth embodiment, after the detector 110 and the digital signal processing unit 120 are properly installed, the gray prediction operation unit 130 and the extension type neural network classifier 140 are constructed, the fuel cell failure prediction system of the present embodiment can be constructed. The operation is as follows: First, as shown in step 201, the detector 110 obtains a plurality of status signals of the fuel cell system 100; then, as shown in step 202, the digital signal processing unit 120 digitizes the status signals and extracts the feature values to form At least one set of characteristic signals is passed to The gray prediction operation unit 130 and the extension type neural network classifier 140. Next, as shown in step 203, the gray prediction operation unit 130 generates a set of gray prediction signals according to the characteristic signals to reflect the status signals that may be generated by the fuel cell system 100 in the next cycle. Of course, at the same time, the above characteristic signal can also be input from the extension type neural network classifier 140 to confirm whether the fuel cell system 100 has a fault in the current cycle. As described above, as shown in step 204, the extension-type neural network classifier 140 analyzes the gray prediction signal using the extension-type neural network to implement a potential fault symptom analysis, thereby generating detection data and prediction data. Finally, as shown in step 205, the detected data and the predicted data are passed to the display 150 to alert the user whether the parts need to be replaced for repair without having to actually fail the fuel cell system 100.

The present invention has been disclosed in the above embodiments, but it is not intended to limit the invention, and it is obvious to those skilled in the art that various modifications and refinements can be made without departing from the spirit and scope of the invention. The scope of protection is subject to the definition of the scope of the patent application attached.

100‧‧‧ fuel cell system

110‧‧‧Detector

111‧‧‧ Pressure detector

112‧‧‧Flow detector

113‧‧‧Battery Detector

120‧‧‧Digital Signal Processing Unit

130‧‧‧ Gray Prediction Unit

140‧‧‧Extreme-like neural network classifier

150‧‧‧ display

160‧‧‧Wireless signal transceiver

170‧‧‧ computer

201-205‧‧‧Steps

The above and other objects, features, advantages and embodiments of the present disclosure will be more apparent and understood. The description of the drawings is as follows: FIG. 1 is a functional block of a fuel cell fault prediction system according to an embodiment of the present disclosure. Figure.

Fig. 2 is a detailed structural view of the fuel cell system 100 of Fig. 1.

Figure 3A is an extension-like neural network classifier 140 of Figure 1 or the like. Schematic diagram of the structure of the neural network.

Figure 3B is a schematic diagram of the extension distance of Figure 3A.

Figure 3C is a schematic diagram of the extension distance before the weight adjustment of Figure 3B.

The 3D diagram is a schematic diagram of the extension distance after weight adjustment in FIG. 3B.

Figure 4 is a flow chart of the learning steps of the extension-like neural network of Figure 3A.

Figure 5 is a flow chart showing the operational steps of the extension-like neural network of Figure 3A.

Figure 6 is a flow chart showing the working steps of the fuel cell failure prediction system of Figure 1.

100. . . Fuel cell system

110. . . Detector

120. . . Digital signal processing unit

130. . . Gray prediction unit

140. . . Extension class neural network classifier

150. . . monitor

Claims (9)

A fuel cell fault prediction system includes: a plurality of detectors for monitoring a plurality of status signals of a fuel cell system; and a digital signal processing unit for generating at least one set of characteristic signals based on the plurality of status signals a gray prediction operation unit for generating a set of gray prediction signals according to the set of characteristic signals; an extension type neural network classifier, which is trained by a plurality of fault models for generating signals according to the set of characteristic signals; Detecting data, and generating a fault prediction data according to the gray prediction signal; a display for displaying the detection data and fault prediction data; and a wireless signal transceiver for wirelessly transmitting the characteristic signal Transfer to the gray prediction operation unit and the extension type neural operation unit. A method for establishing a fuel cell fault prediction system includes: setting a plurality of detectors to monitor a plurality of state signals of a fuel cell system; establishing a fuel cell model by using the state signals; and generating a complex array fault feature by using the fuel cell model Signaling; establishing a plurality of matter-element models according to the fault characteristic signals; using the matter-element model to train an extension-type neural network system; and using the fault feature signals to establish a gray prediction model; The gray prediction model is placed in front of the extension-type neural network system to generate a prediction feature signal to the extension-type neural network system according to the received measured characteristic signal. The method for establishing a fuel cell failure prediction system according to claim 2, wherein the fuel cell model comprises an electrochemical submodel and a thermodynamic submodel. The method for establishing a fuel cell failure prediction system according to claim 3, wherein the electro-chemical sub-model is used to describe a voltage drop generated by an activated anode and a cathode in the fuel cell _Vact : Where ξ _i (i=1-4) is the characteristic coefficient of the fuel cell, and I _FC is the fuel cell current value. (Unit: atmospheric pressure / atm) is the oxygen concentration. The requested item failure prediction system of the fuel cell 3 of the method for establishing, wherein the sub-model-based electrification to describe the ohmic drop within the fuel cell voltage _{V ohmic: V ohmic = - I} FC (R M + R C) wherein, R _C is the resistance of the contact electron flow, and R _M is the impedance through the proton exchange membrane. The method for establishing a fuel cell failure prediction system according to claim 3, wherein the electro-chemical sub-model is used to describe a concentration loss caused by diffusion-limited mass transfer in the fuel cell V _con :V _con =- B ln[1- ( J / J _MAX )] where B (unit: volt / V) is the type constant of the fuel cell, J _MAX is the maximum current density, and J is the current density generated by the fuel cell. The method for establishing a fuel cell failure prediction system according to claim 3, wherein the thermodynamic submodel is used to describe a reversible voltage E _{thermo of a} fuel cell output thermodynamics: among them, Is the pressure of hydrogen required for the fuel cell, It is the pressure of the oxygen required by the fuel cell, and T is the operating temperature of the fuel cell. The method for establishing a fuel cell fault prediction system according to claim 2, wherein the matter element model comprises a fan fault element model, a heat dissipation system fault matter element model, a fuel consumption oversized matter element model, and a hydrogen pressure shortage. The matter element model, a vent hole plugging element model and an overload matter element model. The method for establishing a fuel cell fault prediction system according to claim 8, wherein each of the matter element models includes a hydrogen pressure fault characteristic value, a fuel cell operating temperature characteristic value, a fuel cell voltage characteristic value, and a fuel cell current. The characteristic value, an input air flow characteristic value, and an exhaust port relative humidity characteristic value.