Failure prediction system of fuel battery
The fuel cell failure prediction system uses machine learning and controlled learning conditions to accurately predict failures by comparing estimated and actual anode gas pressure rates, addressing inaccuracies in existing systems due to varying conditions.
Patent Information
- Application Number
- JP2023070779
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-04-24
- Publication Date
- 2024-11-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing fuel cell failure prediction systems struggle to accurately determine whether the measured rate of decrease in anode gas pressure falls within normal variation ranges due to variations caused by aging, temperature, and other conditions, leading to false or overlooked failures.
A fuel cell failure prediction system utilizing a server and vehicles equipped with sensors, communication units, and machine learning to calculate an estimated rate of decrease in anode gas pressure under specific conditions, reducing explanatory variables by limiting learning conditions, and comparing estimated and actual rates to predict failures accurately.
The system achieves high-accuracy prediction of fuel cell failures by using machine learning to account for variations in anode gas pressure, reducing false detections and improving predictive accuracy through controlled learning conditions and real-time comparisons.
Smart Images

Figure 2024156374000001_ABST
Abstract
Description
[Technical field]
[0001] The technology disclosed in this specification relates to a fuel cell failure prediction system. [Background technology]
[0002] The fuel cell failure prediction system disclosed in Patent Document 1 measures the rate at which the anode gas pressure drops while the supply of cathode gas to the fuel cell is ensured, and if the rate is greater than a predetermined threshold rate, it is determined that a cross leak anomaly has occurred in the fuel cell. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] JP 2019-149255 A Summary of the Invention [Problem to be solved by the invention]
[0004] The rate at which the anode gas pressure drops varies greatly depending on conditions such as aging, temperature, and pressure. In this case, the technology of Patent Document 1 may not be able to accurately determine whether the measured rate of drop is within the normal range of variation or an early failure. There is a risk of erroneous detection or overlooking of a failure. [Means for solving the problem]
[0005] The technology disclosed in this specification is embodied in a fuel cell failure prediction system. The failure prediction system includes a server and a plurality of vehicles. Each of the plurality of vehicles includes a fuel cell, a sensor for measuring the anode gas pressure of the fuel cell and a physical value related to the fuel cell, and a vehicle communication unit for transmitting the anode gas pressure and the physical value to a server and receiving a signal from the server. The server includes a server communication unit for receiving the anode gas pressure and the physical value from each of the plurality of vehicles and transmitting a signal to each of the plurality of vehicles. The server includes a learning unit that learns, by machine learning, an estimation formula for calculating an estimated rate of decrease of the anode gas pressure under a specific condition based on the anode gas pressure and the physical value received from the plurality of vehicles. The specific condition is a condition in which a valve that affects the anode gas pressure is closed and there is no load on the fuel cell. The server includes an estimation unit that calculates an estimated rate of decrease of the anode gas pressure based on the estimation formula using a physical value acquired under a specific condition among the physical values received from the plurality of vehicles. The server includes a calculation unit that calculates an actual measured rate of decrease, which is the actual rate of decrease of the anode gas pressure, based on the anode gas pressure received from the plurality of vehicles. The server includes an identification unit that identifies a specific vehicle from among the plurality of vehicles, the specific vehicle being a vehicle having a difference between an estimated decrease speed and an actual decrease speed that is greater than a predetermined threshold value. When the specific vehicle is identified, the server communication unit transmits a signal to the specific vehicle.
[0006] According to this configuration, the estimated rate of decrease of the anode gas pressure can be calculated using machine learning based on physical values collected from a plurality of vehicles. Since the variation in the rate of decrease of the anode gas pressure can be taken into consideration, it is possible to calculate the estimated rate of decrease with high accuracy. Furthermore, by limiting the conditions for learning the estimation formula for calculating the estimated rate of decrease to specific conditions, it is possible to reduce explanatory variables. This also makes it possible to improve the calculation accuracy of the estimated rate of decrease. Then, by comparing the estimated rate of decrease with the actually measured rate of decrease, it is possible to predict failure of the fuel cell with high accuracy. [Brief description of the drawings]
[0007] [Figure 1] 1 is a schematic diagram of a fuel cell failure prediction system 1. [Diagram 2] 1 is a diagram illustrating the configuration of a vehicle 100. [Diagram 3] 6 is a flowchart illustrating a fuel cell failure prediction process performed by the server 50. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0008] (Configuration of fuel cell failure prediction system 1) A fuel cell failure prediction system 1 according to an embodiment will be described with reference to Figure 1. The fuel cell failure prediction system 1 is a system that predicts failures in fuel cells equipped in a plurality of vehicles 100. The vehicles 100 are so-called fuel cell vehicles, and are vehicles that run on roads.
[0009] As shown in Fig. 1, the failure prediction system 1 includes a server 50 and a plurality of vehicles 100. Each vehicle 100 is connected to a communication network CN including wireless communication, such as a mobile communication network. The server 50 is also connected to the communication network CN via a server communication unit 52. Therefore, each vehicle 100 and the server 50 can communicate with each other via the communication network CN.
[0010] The server 50 mainly includes a control unit 51 and a server communication unit 52. The control unit 51 is a device that controls the server 50. The control unit 51 can be configured by, for example, a computer including a CPU, a ROM, a RAM, and the like. When the computer executes a program, the control unit 51 functions as a learning unit 53, an estimation unit 54, a calculation unit 55, a specification unit 56, and the like shown in FIG. 1. The details of these units will be described later.
[0011] (Configuration of vehicle 100) The configuration of the vehicle 100 will be described with reference to Fig. 2. The vehicle 100 mainly includes an ECU 101, a vehicle communication unit 102, and a fuel cell system 103. The fuel cell system 103 mainly includes a fuel tank 110, an injector 120, an ejector 130, a fuel cell stack 140, and a gas-liquid separator 150.
[0012] The fuel tank 110 stores an anode gas (hydrogen gas in this embodiment) to be supplied to the fuel cell stack 140. The fuel tank 110 is connected to an injector 120. The injector 120 adjusts the pressure and flow rate of the anode gas supplied to the fuel cell stack 140. A downstream end of the injector 120 is connected to an ejector 130 and a pressure sensor 124. The pressure sensor 124 detects the anode gas pressure in the fuel supply path. A downstream end of the ejector 130 is connected to the fuel cell stack 140. In addition, a gas circulation path 152 is connected to the ejector 130.
[0013] The fuel cell stack 140 is also connected to an upstream end of an exhaust gas passage 142. The downstream end of the exhaust gas passage 142 is connected to a gas-liquid separator 150. The fuel off-gas is supplied to the gas-liquid separator 150 through the exhaust gas passage 142. A temperature sensor 143 is disposed in the exhaust gas passage 142. The temperature sensor 143 measures the outlet water temperature of the fuel cell stack 140. The gas-liquid separator 150 separates and stores water contained in the fuel off-gas. The gas-liquid separator 150 is connected to an upstream end of a gas circulation passage 152. The fuel off-gas in the gas-liquid separator 150 is supplied to the ejector 130 through the gas circulation passage 152. The gas-liquid separator 150 is also connected to an upstream end of an exhaust drainage passage 156. The downstream end of the exhaust drainage passage 156 is open to the outside. An exhaust drainage valve 158 is provided in the exhaust drainage passage 156.
[0014] In addition, a downstream end of an air supply path 160 is connected to the fuel cell stack 140. An upstream end of the air supply path 160 is open to the outside. A compressor 162 is provided in the air supply path 160. The compressor 162 pressurizes the air introduced into the air supply path 160 to the fuel cell stack 140. In addition, an upstream end of an air discharge path 164 is connected to the fuel cell stack 140.
[0015] The ECU 101 includes a CPU and memories such as a ROM and a RAM. The ECU 101 specifies a load (required load) required of the fuel cell system 103, and controls the operation of the injector 120, etc. so as to obtain a required current.
[0016] Furthermore, the ECU 101 acquires various physical values related to the fuel cell system 103. Examples of the physical values include values indicating the internal state of the fuel cell (outlet water temperature measured by the temperature sensor 143, low-pressure hydrogen pressure, estimated water content), values that are degradation factors of the fuel cell (accumulated mileage, accumulated operating time), etc. The low-pressure hydrogen pressure is the anode gas pressure controlled to a low pressure by the injector 120. The low-pressure hydrogen pressure is, for example, the pressure used during a stop period of the fuel cell system 103. The estimated water content is an estimate of the amount of water contained in the fuel off-gas.
[0017] The vehicle communication unit 102 is a part that communicates with the server 50 via the communication network CN. The vehicle communication unit 102 transmits to the server 50 the low-pressure hydrogen pressure measured by the pressure sensor 124 and various physical values acquired by the ECU 101. The low-pressure hydrogen pressure and the physical values may be time-series data. The vehicle communication unit 102 also receives various signals from the server 50.
[0018] (Fuel cell failure prediction processing) 3, the fuel cell failure prediction process performed by the server 50 will be described. This process is a process for predicting the occurrence of cross leakage, which is one of the failure modes of the fuel cell system 103. In the following, "Step 10" will be abbreviated as "S10".
[0019] In S10, the server communication unit 52 receives the time-series data (low-pressure hydrogen pressure, physical values) transmitted from each of the multiple vehicles 100.
[0020] In step S20, the control unit 51 determines whether or not a predetermined period of time has passed since the last time the estimation equation was learned. The predetermined period of time is a period that can be determined in advance, and may be, for example, several weeks or several months.
[0021] If the predetermined period has elapsed (S20: YES), the process proceeds to S30. In S30, the learning unit 53 learns the estimation formula. A more detailed explanation will be given. First, specific conditions are set as conditions for learning. In this embodiment, the first specific condition is that the valves that affect the low-pressure hydrogen pressure are closed. In other words, the injector 120 and the exhaust drain valve 158 are closed. The second specific condition is that there is no current load on the fuel cell system 103. An example of a situation in which these specific conditions are met is when the ignition switch of the vehicle 100 is off.
[0022] Next, explanatory variables acquired under specific conditions are extracted. Various values related to the fuel cell system 103 can be used as explanatory variables. In this embodiment, the low-pressure hydrogen pressure and various physical values transmitted from the multiple vehicles 100 are used as explanatory variables. Specifically, data points that match the specific conditions are extracted from all time-series data received from the multiple vehicles 100. In other words, data measured during a period in which the specific conditions are met is extracted. Examples of the extracted data include the low-pressure hydrogen pressure, the outlet water temperature, the estimated water content, the accumulated mileage, and the accumulated driving time. The low-pressure hydrogen pressure may be an initial value.
[0023] Finally, an estimation formula with the reduction rate of the low-pressure hydrogen pressure as the objective variable is learned by machine learning using the extracted explanatory variable data. The learned estimation formula is stored in a storage unit (not shown). Then, the process proceeds to S40.
[0024] On the other hand, if the predetermined period has not elapsed (S20: NO), the process proceeds to S25, where the control unit 51 reads out an estimation formula stored in a storage unit (not shown), and then the process proceeds to S40.
[0025] In S40, the control unit 51 selects one vehicle 100 for which failure prediction processing has not been executed from among the multiple vehicles 100 for which time-series data was received in S10. In S50, the control unit 51 judges whether or not there is a data point that matches a specific condition in the time-series data received from the selected vehicle 100. The contents of this judgment have been explained in S30. If there is no data point (S50: NO), skip to S90, and if there is (S50: YES), proceed to S60.
[0026] In S60, the estimation unit 54 extracts data measured during a period in which a specific condition is satisfied. The data extraction has already been described in S30. The estimation unit 54 uses the extracted explanatory variable data to calculate an estimated rate of decrease in the low-pressure hydrogen pressure based on the estimation formula acquired in S25 or S30. Then, based on the estimated rate of decrease, it calculates an estimated value of the amount of decrease in the low-pressure hydrogen pressure (estimated decrease amount).
[0027] In S70, the calculation unit 55 extracts data points that meet specific conditions from the time series data of low-pressure hydrogen pressure received from the selected vehicle 100. Then, based on the extracted low-pressure hydrogen pressure, it calculates an actual decrease rate, which is the actual decrease rate of the low-pressure hydrogen pressure. It also calculates an actual value of the decrease amount of the low-pressure hydrogen pressure (actual decrease amount) based on the actual decrease rate. In S80, the calculation unit 55 calculates the deviation, which is the difference between the estimated decrease amount and the actual decrease amount.
[0028] In S90, the control unit 51 judges whether or not the deviation degree has been calculated for all of the multiple vehicles 100. If the deviation degree has not been calculated for all of the vehicles (S90: NO), the process returns to S40, an unprocessed vehicle is newly selected, and the deviation degree is calculated. On the other hand, if the deviation degree has been calculated for all of the vehicles (S90: YES), the process proceeds to S100.
[0029] In S100, the identification unit 56 determines whether or not there is a specific vehicle among the multiple vehicles 100, the deviation of which is greater than a predetermined threshold. The predetermined threshold is any value that can be determined in advance. A specific vehicle is a vehicle in which the actual deterioration amount deviates significantly from the estimated deterioration amount, and the occurrence of a cross leak is predicted. If there is no specific vehicle (S100: NO), the flow ends, and if there is a specific vehicle (S100: YES), the flow proceeds to S110.
[0030] In S110, the server communication unit 52 transmits a notification signal to the specific vehicle. If there are multiple specific vehicles, the notification signal is transmitted to all of the vehicles. The specific vehicle that receives the notification signal can notify the user of the abnormality by diagnosis (self-diagnosis function). This makes it possible to detect a sign of minor cross leak at the time of its occurrence and notify the user. Before the cross leak worsens and is judged to be a malfunction, resulting in vehicle stoppage, it is possible to reduce downtime by taking action such as repairs.
[0031] (effect) The problem will be explained. As a method for detecting the occurrence of cross leakage in a fuel cell, there is a method for monitoring the amount of decrease in low-pressure hydrogen pressure. For example, a certain threshold (e.g., 50% of the failure judgment threshold) is set, and when the amount of decrease in low-pressure hydrogen pressure exceeds this threshold, the sign is detected. However, the amount of decrease in low-pressure hydrogen pressure is largely dependent on the internal state such as aging deterioration, temperature, and pressure. Therefore, there is a risk of false detection or overlooking of a failure, especially in the initial state of a failure. Therefore, in the technology of this specification, it is possible to calculate the estimated amount of decrease in low-pressure hydrogen pressure using machine learning based on time-series data (low-pressure hydrogen pressure, physical values) collected from multiple vehicles 100 (S60). The variation in the amount of decrease in low-pressure hydrogen pressure due to the internal state and deterioration state of the fuel cell can be taken into consideration using big data. Therefore, it is possible to calculate the estimated amount of decrease with high accuracy. In addition, by limiting the conditions for learning the estimation formula to specific conditions, it is possible to reduce explanatory variables (S30). This also makes it possible to improve the calculation accuracy of the estimated amount of decrease. Then, by comparing the estimated decrease amount with the actually measured decrease amount, a failure in the fuel cell can be predicted with high accuracy (S100).
[0032] In the technology of this specification, the estimation formula can be periodically re-learned (S20, S30). This makes it possible to incorporate the influence of seasonal factors. This makes it possible to predict fuel cell failures with a higher degree of accuracy.
[0033] (Modification) The specific conditions in this embodiment are merely examples, and various conditions can be used as the specific conditions.
[0034] The method of predicting the occurrence of cross leakage is not limited to the method using the deviation of the decrease amount, but may be, for example, a method using a comparison between an estimated decrease rate of the low-pressure hydrogen pressure and an actually measured decrease rate. [Explanation of symbols]
[0035] 1: Failure prediction system 50: Server 51: Control unit 52: Server communication unit 53: Learning unit 54: Estimation unit 55: Calculation unit 56: Identification unit 100: Vehicle 101: ECU 102: Vehicle communication unit 103: Fuel cell system 124: Pressure sensor 140: Fuel cell stack 143: Temperature sensor
Claims
[Claim 1] A fuel cell failure prediction system including a server and a plurality of vehicles, Each of the plurality of vehicles includes: A fuel cell; a sensor for measuring an anode gas pressure of the fuel cell and a physical value related to the fuel cell; a vehicle communication unit that transmits the anode gas pressure and the physical value to the server and receives a signal from the server; It is equipped with The server, a server communication unit that receives the anode gas pressure and the physical value from each of the plurality of vehicles and transmits a signal to each of the plurality of vehicles; a learning unit that learns, by machine learning, an estimation formula for calculating an estimated rate of decrease of the anode gas pressure under specific conditions based on the anode gas pressure and the physical value received from the plurality of vehicles, the specific conditions being a condition in which a valve that affects the anode gas pressure is closed and there is no load on the fuel cell; and an estimation unit that calculates an estimated rate of decrease of the anode gas pressure based on the estimation equation, using the physical values acquired under the specific condition among the physical values received from the plurality of vehicles; a calculation unit that calculates an actual decrease rate of the anode gas pressure, which is an actual decrease rate of the anode gas pressure, based on the anode gas pressures received from the plurality of vehicles; an identification unit that identifies a specific vehicle from the plurality of vehicles, the specific vehicle being a vehicle in which a difference between the estimated decrease speed and the actual decrease speed is greater than a predetermined threshold; Equipped with A fuel cell failure prediction system, wherein the server communication unit transmits the signal to the specific vehicle when the specific vehicle is identified.
Citation Information
Patent Citations
Fuel battery system and control method for fuel battery
JP2019149255A
