System and method for monitoring gas analytes

Abstract

The present disclosure relates to a system and method for monitoring and classifying gases emitted from a closed system having a gas source by means of a gas sensor that is a priori pre-trained to distinguish off-gas events (OGE) or heat effluent events (TRE) from non-OGE interfering gas emissions. Pre-training utilizes one of a machine learning (ML) or deep learning (DL) algorithm to pre-train a gas sensor to detect a plurality of known gas analytes to generate sensor signals having unique characteristics, and extracts features from said sensor signals to establish an estimated probability or decision boundary for false positive emissions of OGE or TRE from non-OGE type interfering gas emissions. The established decision boundary or probability distribution is implemented as a candidate model for a field placement to classify said emitted gases to distinguish whether OGE or TRE has occurred.

Description

Technology Field

[0001] Cross-reference regarding related applications

[0002] This application claims priority and benefits from U.S. provisional patent application serial number 63 / 040,260, titled “System and method for monitoring gas analytes,” filed on June 17, 2020. The present application also refers to U.S. Application No. 15 / 637,381 filed on June 29, 2017, as U.S. Patent 10,877,011B2 granted on December 29, 2020, which claims the benefit of U.S. Provisional Application No. 62 / 356,111 filed on June 29, 2016, titled “System and method for detecting and controlling an analyte” and U.S. Provisional Application No. 62 / 454,516 filed on February 3, 2017, titled “System comprising an energy storage enclosure and monitoring the same”, the contents of which are incorporated herein by reference in their entirety.

[0003] Technology field

[0004] The present disclosure relates to a system and method for monitoring and classifying gases emitted from a closed system having a gas source by means of a gas sensor that is a priori pre-trained to distinguish an off-gas event (OGE) or a thermal run-off event (TRE) from a non-off-gas event (non-OGE). Background Technology

[0005] A battery is an electronic device capable of storing high-density electrical energy. As with any battery, conditions for a thermal runaway event (TRE) can occur during discharge and charging. For example, thermal runaway can be initiated by a short circuit within the battery (e.g., a battery cell), improper battery use, physical abuse, manufacturing defects, or exposure of the battery to extreme external temperatures. Thermal runaway occurs when the battery's internal reaction rate increases to a point where it generates more heat than it can draw out, causing both the internal reaction rate and the generated heat to increase further.

[0006] The impact of thermal runaway conditions can depend on the battery type. For example, in flooded electrolyte batteries such as lead-acid batteries, thermal runaway conditions can cause the electrolyte to boil, resulting in the leakage of dangerous electrolyte gases into the surrounding environment, also known as off-gas events (OGE). In sealed batteries, such as lithium-ion batteries used in devices like electric vehicles, laptops, and mobile phones, thermal runaway conditions can induce expansion, which can cause the sealed battery to explode, release dangerous electrolyte gases into the environment, or pose a fire hazard.

[0007] The present disclosure relates to a system and method for monitoring and classifying gases emitted from a closed system having a gas source by means of a gas sensor that is a priori pre-trained to distinguish off-gas events (OGE) or heat effluent events from non-OGE interfering gas emissions. Pre-training utilizes one of a machine learning (ML) or deep learning (DL) algorithm to pre-train a gas sensor to detect a plurality of known gas analytes to generate sensor signals having unique characteristics, and extracts features from the sensor signal to establish a decision boundary or estimated probability of false positive emissions of OGE or TRE from non-OGE type interfering gas emissions. The established decision boundary or estimated probability may be implemented as a candidate model for field deployment to classify the emitted gas as either OGE or TRE or both in order to distinguish it from non-OGE type interfering gas emissions.

[0008] A method for monitoring and classifying gases emitted from a closed system having a gas source may include the following steps: monitoring a gas source for the emission of a gas analyte by at least one gas sensor having one or more sensing electrodes, wherein the at least one pre-trained gas sensor utilizes one of a machine learning (ML) or deep learning (DL) algorithm a priori before the initial field deployment of the sensor to classify the emitted gas analyte as an event including one or both of an off-gas event (OGE) or a thermal runaway event (TRE) from a non-OGE interference gas emission.

[0009] In the example, utilizing an ML or DL ​​algorithm a priori to pre-train the at least one gas sensor to classify emitted gas analytes comprises at least the following steps: (1) training the at least one gas sensor to detect each and all of the plurality of known gas analytes over a time duration by each of one or more sensing electrodes of the at least one gas sensor to generate a sensor signal representing the unique characteristics of each and all of the plurality of known gas analytes; (2) pre-processing each of the generated sensor signals over the time duration to extract a plurality of corresponding features of each and all of the plurality of known gas analytes; (3) processing the extracted corresponding plurality of features to establish a decision boundary for false positive emissions for one or both of the OGE and TRE, and to establish a decision boundary for each of the non-OGE type interfering gas emissions of each and all of the remainder; and (4) to classify the gas analyte emitted by the gas source as one or both of the OGE or TRE from the non-OGE type of interfering gas emission, the step of storing the set decision boundary in the ML or pre-trained neural network that provides the estimated probability DL algorithm in memory as one or more candidate models for post-field placement of the sensor.

[0010] In another embodiment, a system for monitoring and classifying a gas emitted from a closed system having a gas source may include an enclosure having a gas source; and at least one gas sensor having one or more sensing electrodes placed to monitor the gas source for the emission of a gas analyte, wherein the at least one gas sensor, which is pre-trained prior to placement, a priori utilizes one of a machine learning (ML) or deep learning (DL) algorithm stored as program code in memory for execution by a processor to detect and classify the emitted gas analyte as an event including one or more of an off-gas event (OGE), an interference gas emission event, and a thermal runaway event (TRE).

[0011] In the example, utilizing the ML or DL ​​algorithm a priori to pre-train the at least one gas sensor to classify the emitted gas analytes involves causing the processor to pre-train the at least one gas sensor prior to the initial field deployment of the sensor to perform the following operations, the following operations being: (1) detecting each and all of the plurality of known gas analytes over a time duration by each of one or more sensing electrodes of the at least one gas sensor to generate each sensor signal representing the unique characteristics of each and all of the plurality of known gas analytes; (2) pre-processing each of the generated sensor signals over a time duration to extract corresponding multiple features of each and all of the plurality of known gas analytes; (3) processing the extracted corresponding multiple features to establish a decision boundary for false positive emissions for one or both of the OGE and TRE and to establish a decision boundary for each of the remaining each and all of the non-OGE type interfering gases from the plurality of known gas analytes; and (4) to classify the gas analyte emitted by the gas source as one or more of OGE, interference gas event and TRE, the set decision boundary in the ML or DL ​​algorithm is stored in memory as one or more candidate models for post-field placement of the sensor.

[0012] A disclosed system for monitoring and classifying gases emitted from a closed system having a gas source by means of a gas sensor a priori pre-trained to distinguish between an off-gas event (OGE) or a thermal run-off event (TRE) from non-OGE interference gas emissions may be implemented as a non-transient memory for storing machine-readable instructions. A processor accesses the non-transient memory and executes machine-readable instructions on a machine to perform steps including monitoring a gas source for the emission of a gas analyte by means of at least one gas sensor having one or more sensing electrodes, wherein the at least one pre-trained gas sensor a priori utilizes one of a machine learning (ML) or deep learning (DL) algorithm prior to the initial field deployment of the sensor to classify the emitted gas analyte as an event including one or both of an off-gas event (OGE) or a thermal run-off event (TRE) from non-OGE interference gas emissions.

[0013] In the example, the machine-readable instruction may utilize an ML or DL ​​algorithm to a priori pre-train at least one gas sensor to classify emitted gas analytes, said ML or DL ​​algorithm comprising at least the following steps: (1) training the at least one gas sensor to detect each and all of the plurality of known gas analytes over a time period by each of one or more sensing electrodes of the at least one gas sensor to generate each sensor signal representing the unique characteristics of each and all of the plurality of known gas analytes; a step of pre-processing each of the generated sensor signals over the time period to extract corresponding multiple features of each and all of the plurality of known gas analytes; and a step of processing the extracted corresponding multiple features to set a decision boundary for false positive emissions for one or both of the OGE and TRE, and to set a decision boundary for each of the non-OGE type interfering gas emissions of each and all of the remainder. and to classify the gas analyte emitted by the gas source as one or both of the OGE or TRE from the non-OGE type interfering gas emission, the method may include the step of storing the set decision boundary in the ML or DL ​​algorithm in memory as one or more candidate models for post-field placement of the sensor. Brief explanation of the drawing

[0014] FIG. 1 illustrates an example of a system (100) for monitoring an off-gas event (OGE). FIG. 2 illustrates an example of an equivalent electrical circuit model of a gas sensor that simulates the monitoring and detection of events including off-gas events (OGE), thermal runaway events (TRE), and interference (non-OGE) gas events. Figures 3a and 3b illustrate two separate off-gas events shown as impedance and capacitance changes to be extracted as ML features from the equivalent electrical circuit model of the gas sensor as shown in Figure 2. Figure 4 illustrates an example of OGE for a single-electrode gas sensor from an OGE gas source. Figure 5 illustrates an example of false positive detection from an interfering gas source for a single-electrode gas sensor with a conventional algorithm. Figure 6 illustrates an example of OGE for at least one electrode gas sensor from an OGE gas source. Figure 7 illustrates an example of an interfering gas event for at least one electrode gas sensor from a non-OGE gas source. Figure 8 illustrates an example of a machine learning (ML) classification design process. Figure 9 illustrates an example of a decision boundary separating a real OGE from a false OGE, in which the decision boundary is established by features extracted from a trained ML algorithm. Figure 10 illustrates examples of real and false off-gas events and other gas selectivity at decision boundaries trained with an ML algorithm. Figures 11a and 11b illustrate an exemplary gas sensor pre-training flowchart applying a machine learning (ML) or deep learning (DL) algorithm. Figure 12 illustrates the implementation of a convolutional neural network in deep learning (DL) training. Figure 13 illustrates the implementation of a Long Short-Term Memory (LSTM) neural network in DL training. Figure 14 illustrates a DL framework for time series classification. Figure 15 illustrates an example of a convolutional neural network architecture for performing a DL algorithm. Specific details for implementing the invention

[0015] The present disclosure relates to a system and method for monitoring an enclosed system having a gas source (e.g., a battery) for any one of an off-gas event (OGE), a thermal runaway event (TRE), and an interfering (non-OGE) gas event by means of at least one gas sensor that is a priori pre-trained (i.e., pre-trained at the factory) before initial deployment in the field, so that the generally deployed gas sensor no longer needs to be trained in the field and no reference gas sensor needs to be used to detect gas release events.

[0016] A battery reaching the end of its life may progressively degrade in performance, which may reduce its capacity, cycle life, and safety. A degraded battery may emit a gas that may be referred to as "off-gas." In one example, off-gas may be emitted by the battery during cycling conditions, such as charge and discharge cycles. One or more causes of battery degradation may include improper battery use, physical abuse, manufacturing defects, exposure of the battery to extreme external temperatures, overcharging, etc.

[0017] The system and method described herein can detect off-gas events (OGE) during cycle conditions and provide early warning of thermal runaway event (TRE) conditions. In one example, the early warning may include logic signal output, an audible alarm, a visual alarm, fire suppression, and communication with other systems and users. Off-gas detected during cycle conditions can be interpreted as a warning that the battery is at risk of thermal runaway. By providing early warning, fires, explosions, and injuries caused in response to thermal runaway conditions can be significantly mitigated. Additionally, the system and method described herein can be configured to monitor any type of battery for off-gas conditions. Thus, the system and method described herein can be used to monitor lithium-ion batteries and lead-acid batteries. In broader applications, the system and method described herein can be applied to any enclosed system having a gas source for detecting gas leaks of flammable or toxic gases, for example, in nuclear reactor environments, oil or gas well drilling platforms, coal gas combustion generators, etc.

[0018] The terms “off-gas,” “emitted gas,” and “gas analyte” may be used interchangeably herein and refer to gaseous byproducts of chemical reactions in a gas source, such as a battery. The off-gas (i.e., “emitted gas” and “gas analyte”) may include electrolyte gases, such as volatile electrolyte solvents, volatile components of the electrolyte mixture of a battery, etc. Volatile electrolyte or off-gas analyte species may include at least the following flammable or toxic gases, namely: lithium-ion battery off-gas, dimethyl carbonate, diethyl carbonate, methyl ethyl carbonate, ethylene carbonate, propylene carbonate, vinylene carbonate, carbon dioxide, carbon monoxide, hydrocarbons, methane, ethane, ethylene, propylene, propane, benzene, toluene, hydrogen, oxygen, nitrogen oxides, volatile organic compounds, toxic gases, hydrogen chloride, hydrogen fluoride, hydrogen sulfide, sulfur oxides, ammonia, chlorine, etc. Additionally, the terms "electrode" and "pad" may be used interchangeably to mean an electrically conductive terminal.

[0019] Additionally, the system and method described herein may be composed of a plurality of battery enclosures. Accordingly, the system and method described herein may be used to monitor gas analytes ("off-gas") emitted by one or more batteries located within the battery enclosure. As used herein, the term "battery enclosure" refers to any housing capable of partially encapsulating one or more batteries. In an example, the battery enclosure may include ventilated and non-ventilated battery enclosures. A ventilated battery enclosure may include a ventilation system that may include intake and exhaust. In another example, the battery enclosure may include a battery shipping container.

[0020] Additionally, the term “processor” as used herein may refer to any device capable of executing machine-readable instructions, such as a computer, controller, integrated circuit (IC), microchip, or any other device capable of implementing logic. The term “memory” as used herein may refer to a non-transient computer storage medium, such as volatile memory (e.g., random access memory), non-volatile memory (e.g., hard disk drive, solid-state drive, flash memory, etc.), or a combination thereof.

[0021] FIG. 1 illustrates an example of a system (100) for monitoring an off-gas event (OGE). The system (100) includes at least a gas source, such as a battery (102), and at least a gas sensor (104) positioned to monitor the emission of a gas analyte (102a) from the battery (102). The battery (102) may be a lithium-ion battery having an enclosure (i.e., encapsulated in a protective case). In an embodiment, the gas sensor (104) may be a semiconductor-type gas sensor or any suitable gas sensor for detecting the gas analyte (102a) (e.g., chemical vapor) emitted from the battery (102).

[0022] Unlike most other systems, the deployed gas sensor (104) eliminates the need for the system (100) to use a separate reference sensor to calculate a moving average from a real-time sensor signal (104a) for the detection of an off-gas event (OGE) from the battery (102). In an implementation, the gas sensor (104) may be a sensor that has one or more sensing electrodes (104b) and is pre-trained a priori (e.g., during manufacturing) using one of a machine learning (ML) or deep learning (DL) algorithm (108a) (program code), and said machine learning (ML) or deep learning (DL) algorithm is stored in memory (108) to be executed by a processor (106) to enable the gas sensor (104) to detect and classify any emitted gas analyte (102a) in real time as an event including one or more of an off-gas event (OGE), a thermal runaway event (TRE), and an interfering gas emission event (i.e., non-OGE). In one example, the gas sensor (104), processor (106), and memory (108) may be an integrated chip (101), such as an ASIC semiconductor chip. In another example, the gas sensor (104), processor (106), and memory (108) may each be individual components electrically connected via a wiring harness or mounted on a printed circuit board (PCB).

[0023] A pre-trained gas sensor (104) may store an ML or DL ​​algorithm (108a) as a candidate model in memory (108) to distinguish a sensor signal (104a) detected by the gas sensor (104) as either OGE or TRE from a non-OGE interfering gas event, without any need for a reference gas sensor or any additional need to retrain the gas sensor (104) once deployed in the field.

[0024] Machine learning and training in the algorithm (108a) step may be performed a priori in the factory during the manufacturing process, or may be performed offline at any time before the sensor (104) is physically commissioned or installed in the system (100). Once the sensor (104) is commissioned in the system (100), real-time adaptation is not required. In another option, the ML or DL ​​algorithm (108a) may be retrained or updated by the sensor (104) to learn new encounters with other gas analytes that are not pre-retrained or listed in the database. The goal of this pre-training using the ML or DL ​​algorithm is to be able to detect OGE as well as identify other gas sources detected by the sensor (104), thereby eliminating the need for a reference sensor.

[0025] In the example, a priori pre-training using an ML or DL ​​algorithm (108a) for at least one gas sensor (104) may include at least the following steps: (1) training at least one gas sensor to detect each and all of the plurality of known gas analytes (i.e., training gases) over a time period by each of one or more sensing electrodes (104a) of the at least one gas sensor (104) to generate each sensor signal (104a) representing the unique characteristics of each and all of the plurality of known gas analytes in order to establish a database; (2) pre-processing each of the generated sensor signals (104a) over the time period to extract corresponding multiple features (e.g., impedance and capacitance, see FIG. 3a, 3b through 7) of each and all of the plurality of known gas analytes (102a); (3) a step of processing the extracted corresponding plurality of features to set a decision boundary for false positive emissions for one or both of OGE and TRE (see FIG. 9 and 10), and to set a decision boundary for each of the remaining non-OGE type interfering gas emissions; and (4) a step of storing the decision boundary set in the ML or DL ​​algorithm (108a) as one or more candidate models for post-field placement of the sensor in memory (108) to classify a gas analyte (102a) emitted by a gas source (e.g., battery (102)) as one or both of OGE or TRE from non-OGE interfering gas emissions. An output signal (110), such as an alarm or logic signal, may be transmitted to warn or display on a screen to take preventive measures against causing an accident or damage to the system (100).

[0026] In the example, the gas source may be a rechargeable lithium-ion battery system or an electric energy storage system (102), wherein the gas analyte emitted from the OGE or TRE may be at least one of the following flammable or toxic gases, namely, lithium-ion battery off-gas, dimethyl carbonate, diethyl carbonate, methyl ethyl carbonate, ethylene carbonate, propylene carbonate, vinylene carbonate, carbon dioxide, carbon monoxide, hydrocarbons, methane, ethane, ethylene, propylene, propane, benzene, toluene, hydrogen, oxygen, nitrogen oxides, volatile organic compounds, toxic gases, hydrogen chloride, hydrogen fluoride, hydrogen sulfide, sulfur oxides, ammonia, and chlorine, or a combination thereof.

[0027] In practice, the ML or DL ​​learning method during pre-training may be based on an equivalent circuit model (200) of the sensor (104) as illustrated in FIG. 2. For example, FIG. 2 may represent an equivalent circuit model of a semiconductor gas sensor. However, other gas sensors may use different equivalent circuit models to be represented by different transfer functions. The exemplified gas sensor types and equivalent circuit model examples are not limited.

[0028] The equivalent impedance circuit model (200) of FIG. 2 shows that each gas sensor signal generated by each of one or more sensing electrodes (104b) of at least one gas sensor (104) may include an impedance value based on a first parallel resistor Rn and capacitor CPEvr pair (204) cascaded in series with a second parallel resistor Rs and capacitor CPE pair (206). The first parallel resistor and capacitor pair (204) of the equivalent impedance circuit model (200) can simulate the dynamics of at least one gas sensor response (104a) that may be exposed to a combination of different emitted analyte gases (102a). The impedance equivalent circuit model (200) of the sensor (104) further includes a contact resistor Rc cascaded in series with the first parallel resistor Rn and capacitor CPEvr pair (204).

[0029] Given a gas sensor (104) having an output impedance R and a fixed input voltage V in the LaPlace region(s), the corresponding input / output transfer function can be developed for an equivalent circuit model (200), which can be expressed as follows:

[0030]

[0031] Here, Rc is the contact resistance, Rn and Rs are the resistances of the circuit model, and CPEvr and CPEs are the capacitances of the gains in the circuit model.

[0032] Changes in impedance values ​​(i.e., resistance and capacitance) over time from electrical elements (Rc, Rn, Rs, CPEvr, CPEs) caused by each and all of the multiple known gas analytes (102a) of the circuit model (see FIG. 3a, b through FIG. 7) can be transmitted as sensor signals (104a) for training, which represent the unique characteristics or properties of the detected known gas analytes. An ML or DL ​​learning method simulates the monitoring and detection of the detected gas analytes (102a) emitted from a gas source (i.e., battery (102)) to classify the gas emission event as one or a combination of off-gas events (OGE), thermal runaway events (TRE), and interference (non-OGE) gas events. Known gas analytes may be gases mentioned in training OGE or TRE, and these gases include, but are not limited to, at least the following flammable or toxic gases: lithium-ion battery off-gas, dimethyl carbonate, diethyl carbonate, methyl ethyl carbonate, ethylene carbonate, propylene carbonate, vinylene carbonate, carbon dioxide, carbon monoxide, hydrocarbons, methane, ethane, ethylene, propylene, propane, benzene, toluene, hydrogen, oxygen, nitrogen oxides, volatile organic compounds, toxic gases, hydrogen chloride, hydrogen fluoride, hydrogen sulfide, sulfur oxides, ammonia, chlorine, etc.

[0033] In one example, a mathematical expression using estimation techniques such as the least squares method and gradient algorithms may be used to extract equivalent resistance and capacitance from an equivalent model (200) of a gas sensor (104). In practice, when detecting the emission of a gas analyte (102a), pre-training of an ML or DL ​​algorithm of at least one gas sensor (104) may include distinguishing changes in sensor impedance caused by environmental disturbances caused by one or more of other gases that affect temperature changes, relative humidity, and the partial pressure of oxygen in the environment, leading to the reporting of false positives.

[0034] In one example, the ML algorithm pre-training of at least one gas sensor (104) in the extraction of corresponding multiple features of each and all of a plurality of known gas analytes (see FIG. 8) may include utilizing any one or a combination of features including moving average calculation, Bollinger band, minimum electrode impedance, maximum impedance change rate, maximum impedance recovery rate for each of at least one electrode (104b) of at least one gas sensor (104), and principal component analysis (PCA) linear discriminant analysis, wherein the DL algorithm pre-training the at least one gas sensor in the extraction of corresponding multiple features of each and all of a plurality of known gas analytes is internally included in a hidden layer of a DL neural network (see FIG. 11a, 11b and 12).

[0035] In another example, an ML or DL ​​algorithm that pre-trains at least one gas sensor in setting the decision boundary for false positive emissions for OGE or TRE and the decision boundary for each of the remaining and all non-OGE type interfering gas emissions may include evaluating a sensor signal (104a) generated using any of the decision methods including support vector machines, discriminant analysis or nearest neighbor algorithms, naive Bayes and neural neighbors, linear regression, GLM, support vector regression, GPR, ensemble methods, decision trees, and DL neural networks, wherein the DL neural network includes at least one of a convolutional neural network (CNN) (see FIG. 12) and a Long Short Term Memory (LSTM) network (see FIG. 11a, 11b).

[0036] FIGS. 3a and 3b illustrate two separate off-gas events from a received sensor signal (104a) illustrated as changes in impedance and capacitance over a time duration (e.g., a sudden change from region 320A to region 330A in FIG. 3a, and a sudden change from region 320B to region 330B in FIG. 3b) to be extracted as ML features from an equivalent electrical circuit model of the gas sensor (104) as illustrated in FIG. 2. These data may be training parameters for ML and DL machine learning algorithms for setting decision boundaries (see FIGS. 9 and 10).

[0037] FIG. 4 illustrates an example of OGE for a single-electrode gas sensor (104) from an OGE gas source. In the example, the sensor signal (104a) parameters may include at least a change over time duration in the value of the impedance and moving average that may cause the generation of an alarm in the output signal (110). Likewise, FIG. 5 illustrates an example of false positive detection from an interfering gas source (non-OGE) for a single-electrode gas sensor having a conventional algorithm.

[0038] Figure 6 illustrates an example of OGE for at least one electrode gas sensor from an OGE gas source. Figure 7 illustrates an example of an interfering gas event for at least one electrode gas sensor from a non-OGE gas source. The unique response of each electrode of the gas sensor provides an ML learning algorithm with the ability to distinguish between OGE and common interfering gases in an energy storage facility.

[0039] The ML algorithm has a unique selectivity that allows it to identify not only OGE but also other interfering gases present to enable other diagnostic capabilities. By comparison, an increase in the number of electrodes (104b) of the gas sensor (104) provides a more unique response, which can be a unique characteristic or attribute (e.g., fingerprint) that helps identify the detected gas analytes. Likewise, features extracted from one or more sensing electrodes can establish a dynamic response of one or more dimensions in more accurately setting the false positive determination boundaries of OGE, TRE, and non-OGE type interfering gas events, thereby improving reliability and accuracy in the trained candidate model (by reducing the probability of error) as well as in classifying events and identifying the type of gas analyte.

[0040] FIG. 8 illustrates an example of a machine learning (ML) classification design process or a method by which an ML algorithm can be developed. A pre-training (supervised learning) approach can combine many signal features (from the gas sensor (104)) using both empirical and physical-based impedance information during the data pre-processing phase of the algorithm development phase. This may also include environmental measurements such as temperature, pressure, and relative humidity included in the sensor set provision. For example, signal features may include moving averages, Bollinger bands, minimum electrode impedance, maximum impedance rate of change, maximum impedance recovery rate for each electrode, principal component analysis (PCA), and linear discriminant analysis. Additionally, pre-training (supervised learning) of gas sensors to distinguish OGE or TRE from non-OGE using other techniques may include, but is not limited to, support vector machines, discriminant analysis or nearest neighbor algorithms, basic statistics of time series value distributions (e.g., location, spread, Gaussianity, outlier properties), linear correlations (e.g., autocorrelation, features of the power spectrum), stationarity (e.g., sliding window measurements, prediction errors), information theory, and entropy / complexity.

[0041] Table 1 below provides some examples of different features extracted from the sensor signals (104a) of each of the multiple electrodes (104b), which can be utilized in ML or DL ​​algorithms (108a) for a priori training and to set a decision boundary for classifying gas analytes (102a) as one of OGE, TRE, and non-OGE interfering gas events (illustrated in the process of FIG. 8 and 11a, 11b). Machine learning is a supervised learning technique that can utilize linear regression, nearest neighbor, support vector regression, and neural networks in training to set candidate models (854) to be used when deploying gas sensors (104) in the field.

[0042]

[0043] Table 1 shows some examples of features extracted from each of at least one electrode for each gas type mentioned in the previous section. The transformed data, along with various features extracted from raw data, is written in Table 1 and is used to generate candidate models using an optimization process that retrieves model parameters, including the use of fitting data, evaluates them using test data not used for fitting, and tunes the candidate models until optimal performance can be achieved. Feature extraction in deep learning can be integrated inside a neural network (see Figures 11a, 11b, 12, 13 and 14a and 14b) and does not need to be performed before classification.

[0044] In another example, as part of a pre-training process, features extracted from known gas analytes can be used to train ML or DL ​​algorithms to sufficiently quantify the approximate percentage % or ppm (parts per million) of each identified gas analyte in the detected gas analyte composition, which can be useful in helping to classify one or a combination of OGE, TRE, and non-OGE as interfering gas emissions from a gas source. Table 2 below illustrates some examples of the detected gas analyte compositions.

[0045]

[0046] FIG. 9 illustrates an example of a decision boundary (930) separating a real OGE (910) from a false OGE (920), which is set by features extracted from an ML algorithm (108a) trained over a historical time duration. This implies that OGE was detected within the region (910) and non-OGE was detected within the region (920). In practice, the decision boundary (930) will be of high dimensionality (i.e., much higher than 3 dimensions) based on the time series data, which makes it nearly impossible to plot all decision boundary plots by the example illustrated in FIG. 9 and FIG. 10.

[0047] FIG. 10 illustrates an example of selectivity for real OGE (1010), false OGE, and other gases (1021–1029) at a decision boundary (1030) trained by an ML algorithm (108a). The ML or DL ​​algorithm (108a) may include a selectivity algorithm that uses the same idea of ​​training decision boundaries using a similar process or technique as illustrated in FIG. 8, except that FIG. 8 may show a composite of multiple dimensions corresponding to different extracted features of the plot to show different decision boundaries for each possible interfering gas that was detected. The idea behind using two techniques (i.e., composite dimensions) is to provide the user with some diagnostic information by identifying the detected gases, as well as to make it easier to detect false positives. Other regions (1021–1029) show how different types of interfering gases can be detected through a pre-trained gas sensor (104). FIG. 10 shows all of the actual OGE region (1010) versus multiple false positive (non-OGE) regions (1021-1029), including the decision boundary that enables the ML or DL ​​algorithm to identify false positive interfering gases.

[0048] FIGS. 11a and 11b illustrate an exemplary flowchart for pre-training a gas sensor (104) with a plurality of known gas analytes using a machine learning (ML) algorithm. For example, in step 1102, for each known gas analyte, raw signals (104a), such as resistance and capacitance, may be generated from multiple electrodes (104b) of the gas sensor (104) and may be transmitted to a processor for feature extraction (e.g., changes in impedance or transfer function over a time duration) in step 1104. The feature extraction step (1104) may include a time-to-frequency transformation (such as a Discrete Cosine Transform DCT or Discrete Fourier Transform DFT) to convert the time-domain analog signal into a frequency-domain signal. In step 1106, the extracted features may be organized accordingly. In step 1108, a machine learning (ML) algorithm is applied to the organized data to establish a candidate model (1110) (e.g., a multidimensional decision boundary configuration). By repeating steps 1102 through 1110 for the remainder of a plurality of known gas analytes, the candidate model (1110) can be updated (see step 1111) to establish a database or construct a composite decision boundary plot to complete the training of the ML algorithm (108a), and said ML algorithm (108a) is stored in memory (108) to be executed by the processor (106). FIG. 11b shows details of the ML training step (1108) of FIG. 11a. More specifically, step 1108 can be achieved by repeating the training steps (1108a through 1108n). Each training step (e.g., 1108a) may include sequentially performing operations during convolution, rectified linear unit (ReLU), and pooling operations. The batch model (1112) will be a field-ready ML algorithm when working with a multi-electrode gas sensor (104) to perform gas analyte classification.

[0049] Likewise, such desired classification can be achieved by a deep learning (DL) algorithm utilizing a pre-trained convolutional neural network (e.g., a convolutional neural network CNN and Long Short-Term Memory (LSTM)) and automatic signal feature extraction as shown in FIGS. 12 and 13, as illustrated in FIGS. 11a and 11b.

[0050] Figure 14 illustrates an example of a DL framework for time series classification. Figure 15 illustrates an example of a convolutional neural network architecture for performing pre-training of a DL algorithm using extracted features. A deep learning (DL) algorithm may consist of multiple layers (see Figures 14 and 15) implementing non-linear functions. Each layer may receive input from the output of the previous layer and apply a non-linear transformation (i.e., sequential pipeline processing technique) to compute the output. This non-linear transformation may be determined by trainable parameters in the fitting process. Some deep learning architectures that can be implemented may include convolutional neural networks (see Figure 11b), start-time and eco-state networks.

[0051] Although the specific examples above have been illustrated and described herein, it should be understood that various other changes and modifications may be made without departing from the spirit and scope of the claimed subject matter. Furthermore, although various aspects of the claimed subject matter have been described herein, these aspects do not need to be used in combination. Accordingly, the appended claims are intended to cover all such changes and modifications within the scope of the claimed subject matter.

Claims

Claim 1 A method comprising the step of monitoring at least one gas analyte emitted by a gas source using at least one gas sensor having one or more sensing electrodes, wherein the at least one gas sensor is pre-trained a priori by utilizing one of a machine learning (ML) or deep learning (DL) algorithm prior to the initial field placement of the at least one gas sensor to detect and classify the at least one gas analyte emitted by the gas source as representing an event including one or both of an off-gas event (OGE) or a thermal runaway event (TRE), and wherein the OGE and the TRE are distinguished from non-OGE interfering gas emission events; wherein the a priori by utilizing the ML or DL ​​algorithm to pre-train the at least one gas sensor to classify the at least one gas analyte emitted by the gas source comprises training the at least one gas sensor to detect each of the plurality of known gas analytes potentially emitted by the gas source over a time period by each of the one or more sensing electrodes of the at least one gas sensor to generate a respective sensor signal representing each of the unique characteristics of each of the plurality of known gas analytes; Preprocessing each generated sensor signal detected over the time period to extract corresponding multiple features from each generated sensor signal for each of the plurality of known gas analytes; processing the extracted corresponding multiple features to set a decision boundary indicating a false positive detection for one or both of the OGE and TRE, and to set a decision boundary indicating a false positive detection for each of the remaining non-OGE interfering gas emission events.and a method comprising storing in memory as one or more candidate models for field placement of the at least one gas sensor the set decision boundary in the ML or DL ​​algorithm to classify the at least one gas analyte emitted by the gas source as one or both of the OGE or TRE—where the OGE and the TRE are distinguished from the non-OGE interfering gas emission event.; Claim 2 A method according to claim 1, wherein the gas source comprises a rechargeable lithium-ion battery system or an electric energy storage system, and the OGE or TRE comprises detecting the emission of at least one or a combination thereof of flammable or toxic gases: lithium-ion battery off-gas, dimethyl carbonate, diethyl carbonate, methyl ethyl carbonate, ethylene carbonate, propylene carbonate, vinylene carbonate, carbon dioxide, carbon monoxide, hydrocarbons, methane, ethane, ethylene, propylene, propane, benzene, toluene, hydrogen, oxygen, nitrogen oxides, volatile organic compounds, toxic gases, hydrogen chloride, hydrogen fluoride, hydrogen sulfide, sulfur oxides, ammonia, and chlorine. Claim 3 A method according to claim 1, wherein each gas sensor signal generated by each of the one or more sensing electrodes of the at least one gas sensor comprises an impedance value based on an equivalent impedance circuit model having a first parallel resistor and capacitor pair cascaded in series with a second parallel resistor and capacitor pair. Claim 4 A method according to paragraph 3, wherein the first parallel resistor and the capacitor pair of the equivalent impedance circuit model simulate the dynamics of at least one gas sensor response when exposed to a combination of different emitted analyte gases. Claim 5 In claim 4, the equivalent impedance circuit model of the sensor further comprises a contact resistance cascaded in series with respect to the first parallel resistor and capacitor pair. Claim 6 delete Claim 7 A method according to claim 1, wherein detecting the emission of the gas analyte from the gas source by at least one gas sensor pre-trained by the ML or DL ​​algorithm is performed without using a reference sensor. Claim 8 A method according to claim 1, wherein the ML or DL ​​algorithm for pre-training the at least one gas sensor upon detection of the emission of the gas analyte further comprises distinguishing changes in sensor impedance caused by environmental disturbances, such as changes in temperature, changes in relative humidity, and one or more of other gases affecting the partial pressure of oxygen in the environment, which cause false positive detection of one or both of the OGE and TRE. Claim 9 In claim 1, the ML algorithm for pre-training the at least one gas sensor when extracting each corresponding multiple feature of the plurality of known gas analytes comprises utilizing any one or a combination thereof of features including moving average calculation, Bollinger bands, minimum electrode impedance, maximum impedance change rate, maximum impedance recovery rate for each of at least one electrode on the at least one gas sensor, principal component analysis (PCA), and linear discriminant analysis, and the DL algorithm for pre-training the at least one gas sensor when extracting each corresponding multiple feature of the plurality of known gas analytes comprises being internally included in a hidden layer of a neural network. Claim 10 The method of claim 1, wherein the ML or DL ​​algorithm for pre-training the at least one gas sensor when setting the decision boundary for false positive emissions for the OGE or TRE and setting the respective decision boundary for each of the remaining non-OGE type interfering gas emissions comprises evaluating the generated sensor signal using any one of decision methods including support vector machines, discriminant analysis or nearest neighbor algorithms, Naive Bayes and Neural Neighbor, linear regression, GLM, support vector regression, GPR, ensemble methods, decision trees, and DL neural networks, wherein the DL neural network comprises at least one of Convolutional Neural Networks (CNN), Inception Time Architecture, Eco State Networks, and Long Short Term Memory (LSTM) networks. Claim 11 In the system, an enclosure having a gas source; and at least one gas sensor having one or more sensing electrodes positioned to monitor at least one gas analyte emitted by the gas source—the at least one gas sensor is pre-trained prior to said position by utilizing a machine learning (ML) or deep learning (DL) algorithm stored as program code in memory for execution by a processor to detect and classify the at least one gas analyte emitted by the gas source as representing an event including one or both of an off-gas event (OGE) or a thermal runaway event (TRE), wherein the OGE and the TRE are distinguished from non-OGE interfering gas emission events—and utilizing the ML or DL ​​algorithm a priori to pre-train the at least one gas sensor to classify the at least one gas analyte emitted by the gas source is to detect each of the plurality of known gas analytes that may potentially be emitted by the gas source over a time period by each of the one or more sensing electrodes of the at least one gas sensor to generate a respective sensor signal representing each of the unique characteristics of each of the plurality of known gas analytes; To preprocess each generated sensor signal detected over the time period to extract corresponding multiple features from each generated sensor signal for each of the plurality of known gas analytes; to process the extracted corresponding multiple features to set a decision boundary indicating a false positive detection for one or both of the OGE and TRE, and to set a decision boundary indicating a false positive detection for each of the remaining non-OGE interfering gas emission events;A system that causes the processor to pre-train the at least one gas sensor prior to the initial field placement of the sensor, so as to store in memory the set decision boundary in the ML or DL ​​algorithm as one or more candidate models for the field placement of the at least one gas sensor, in order to classify the at least one gas analyte emitted by the gas source as one or both of the OGE or TRE—where the OGE and the TRE are distinguished from the non-OGE interfering gas emission event. Claim 12 A system according to claim 11, wherein each gas sensor signal generated by each of the one or more sensing electrodes of the at least one gas sensor comprises an impedance value based on an equivalent impedance circuit model having a first parallel resistor and capacitor pair cascaded in series with a second parallel resistor and capacitor pair, and the first parallel resistor and capacitor pair in the equivalent impedance circuit model simulates the dynamic characteristics of the at least one gas sensor response when exposed to different emitting gases. Claim 13 In claim 12, the system for detecting the emission of the gas analyte from the gas source by at least one gas sensor pre-trained by the ML or DL ​​algorithm is performed without using a reference sensor. Claim 14 A system according to claim 11, wherein the ML or DL ​​algorithm for pre-training the at least one gas sensor when extracting each corresponding feature of the plurality of known gas analytes comprises utilizing any one or a combination thereof of moving average calculation, Bollinger bands, minimum electrode impedance, maximum impedance change rate, maximum impedance recovery rate for each of at least one electrode on the at least one gas sensor, principal component analysis (PCA), and linear discriminant analysis. Claim 15 A system according to claim 11, wherein the ML or DL ​​algorithm for setting the decision boundary for false positive emissions for the OGE or TRE and for setting the decision boundary for each of the remaining non-OGE type interfering gas emissions comprises evaluating the generated sensor signal by utilizing any one of decision methods including support vector machine, discriminant analysis or nearest neighbor algorithm, naive Bayes and neural neighbor, linear regression, GLM, support vector regression, GPR, ensemble method, decision tree, and DL neural network, wherein the DL neural network comprises at least one of a convolutional neural network (CNN), a start-time architecture, an echo-state network, and a long-short-term memory (LSTM) network.

Images