METHOD FOR SORTING A SET OF BATTERIES BASED ON A MACHINE CLASSIFICATION MODEL

Abstract

A computer-implemented method for the machine learning of a model for classifying batteries into two categories: functional or defective, the method comprising the following steps: acquiring, on a set of batteries of the same type, a group of measurements characteristic of the operation of a battery, carrying out complete cycling of each battery of the set and measuring at least one curve from among a charge curve or a discharge curve of the battery, determining, for each battery, a label of belonging to the functional or defective category by comparing at least one measured curve with a reference curve characterizing the correct operation of the battery, and carrying out supervised training of a model for classifying a battery according to the two categories based on the groups of measurements and on the label of belonging to one of the two categories.

Claims

1. A computer-implemented method for the machine learning of a model for classifying batteries into two categories: functional or defective, the method comprising the following steps: acquiring, on a set of batteries of the same type, a group of measurements characteristic of the operation of a battery, carrying out complete cycling of each battery of said set and measuring at least one curve from among a charge curve or a discharge curve of the battery, determining, for each battery, a label of belonging to the functional or defective category by comparing at least one measured curve with a reference curve characterizing the correct operation of the battery, and carrying out supervised training of a model for classifying a battery according to the two categories based on said groups of measurements and on said label of belonging to one of the two categories.

2. The machine learning method according to claim 1, wherein the supervised training of the classification model is carried out by way of a random forest algorithm.

3. The machine learning method according to claim 1, wherein the group of measurements characteristic of the operation of a battery comprises at least one measurement of an open-circuit voltage across the terminals of the battery, for example an average measurement over a predetermined time interval.

4. The machine learning method according to claim 1, wherein the group of measurements characteristic of the operation of a battery comprises at least one measurement of a profile of a current flowing through the battery during a partial charge.

5. The machine learning method according to claim 4, wherein the current profile is measured at at least one particular point from among a maximum current value and an end-of-partial-charge current value.

6. The machine learning method according to claim 1, wherein the group of measurements characteristic of the operation of a battery comprises at least one measurement of a profile of a voltage across the terminals of the battery during relaxation of the battery following a partial charge.

7. The machine learning method according to claim 6, wherein the voltage profile measurement comprises at least one estimate, performed through linear regression, of the slope of the voltage curve during discharging and/or measuring an end-of-relaxation potential.

8. The machine learning method according to claim 1, wherein the group of measurements characteristic of the operation of a battery comprises at least one electrochemical impedance spectroscopy measurement carried out by exciting the battery with a signal at at least one predetermined frequency.

9. The machine learning method according to claim 8, wherein the electrochemical impedance spectroscopy measurement is carried out for at least two frequencies predetermined at 1000 Hz and 10 Hz.

10. The machine learning method according to claim 1, wherein the group of measurements characteristic of the operation of a battery comprises at least one measurement of the position of a battery on a fabrication wafer.

11. The machine learning method according to claim 1, wherein the batteries are Li-free solid batteries.

12. A computer-implemented classification model obtained using the machine learning method according to claim 1.

13. A method for testing a set of batteries comprising the following steps, for each battery: acquiring the same group of measurements used to train the classification model according to claim 12, and implementing the classification model according to claim 12 to determine, based on said group of measurements, whether the battery belongs to the functional or defective category.

14. A non-transitory computer-readable storage medium storing a computer program comprising code instructions for implementing the method according to claim 1.

15. A device for testing a set of batteries, comprising a measurement apparatus (Vn, An) and a computing device (Cn) that are configured together to carry out the test method according to claim 13.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Other features and advantages of the present invention will become more apparent on reading the following description with reference to the following appended drawings.

[0050] FIG. 1 shows a general diagram of the two-phase principle of the test method according to the invention,

[0051] FIG. 2 shows a graph of a measurement of an open-circuit voltage across the terminals of a microbattery,

[0052] FIG. 3 shows a graph of a current profile during partial charging of a microbattery,

[0053] FIG. 4 shows a graph of a potential relaxation profile of a partially charged microbattery,

[0054] FIG. 5 shows a Nyquist plot of an electrochemical impedance of a microbattery as measured using spectroscopy for various frequencies,

[0055] FIG. 6 shows a database of measurements used at input of a learning method according to the invention,

[0056] FIG. 7 shows a set of reference curves characterizing a functional microbattery and multiple examples of curves characterizing a defective microbattery,

[0057] FIG. 8 illustrates a step of constructing a node during the generation of a tree via a random forest learning algorithm,

[0058] FIG. 9 shows one example of a tree generated by a random forest learning algorithm,

[0059] FIG. 10 shows two examples of classifying microbatteries produced on a silicon wafer and tested by way of the machine classification method according to the invention, and

[0060] FIG. 11 shows a diagram of a device for testing batteries according to one embodiment of the invention.

DETAILED DESCRIPTION

[0061] FIG. 1 schematically shows a flowchart detailing the steps for implementing the invention, which comprises two separate phases: a first phase of learning 101 the classification model, carried out on a sample of multiple batteries, and a second phase of using 102 the trained and validated model to classify new manufactured batteries into two categories: functional battery or defective battery.

[0062] The learning phase 101 comprises a first step 110 of acquiring data, through measurements carried out on each battery, and then a step of completely cycling 111 the batteries, comprising at least one complete charge and/or one complete discharge of the battery. The complete cycling step 111 also comprises a charge and/or discharge profile measurement in order to characterize the battery as belonging to a functional or defective category. Based on the measurements acquired in step 110 and on the labelling into two categories carried out in step 111, a supervised machine classification model is trained, in step 112, to learn to classify a battery as functional or defective based on just the measurements carried out in step 110.

[0063] Once the classification model has been trained and validated, it is used, in the second phase 102, in order to be applied 113 directly to measurements carried out in step 110 on new manufactured batteries in order to classify them directly (step 114) into the functional or defective categories without having to completely cycle the battery.

[0064] Each of the steps of the method is now detailed.

[0065] Step 110 of collecting measurements of parameters representative of the correct operation of a battery is carried out both for the learning phase 101, on a given set of batteries, and in the use phase 102. One exemplary embodiment of step 110 that relates more particularly to microbatteries, in particular “Li-free” microbatteries, is detailed. Without departing from the scope of the invention, other parameters may be envisaged for characterizing the correct operation of a battery, in particular of another type.

[0066] For microbatteries in particular, the following measurements are envisaged to characterize the operating state thereof.

[0067] A first type of measurement relates to a measurement of the open-circuit voltage or OCV. This voltage is measured for a time interval, for example equal to a few seconds, as soon as the battery leaves the manufacturing process, before any electrical loading. The measured voltage values may be averaged over the measurement time interval. This voltage is measured directly across the terminals of the battery.

[0068] FIG. 2 illustrates one example of an open-circuit voltage measurement curve.

[0069] Another type of measurement relates to the current profile during a partial charge of the battery. For this purpose, a partial charge is initiated for a time interval, for example equal to 1 minute, by applying a predetermined voltage (for example equal to 4.2 V) across the terminals of the battery. During this time interval, the current flowing through the battery is then measured, and this gives a current evolution curve of the type shown in FIG. 3. This current profile may be used in its entirety or else certain characteristic points may be extracted therefrom, for example the maximum value of the current Imax or the end-of-charge current value Ip. Other points of the curve may also be envisaged.

[0070] After the partial charging of the battery, the loading is stopped and the potential relaxation of the loaded battery may be ascertained by measuring the voltage across the terminals of the battery again for a given time interval. One example of such a measurement is shown in FIG. 4, with the complete curve on the left and a zoom-in on the amplitude of this curve on the right.

[0071] In this case too, certain characteristic points or parameters may be selected, such as the slope a_relax of the relaxation curve and the end-of-relaxation potential Vr.

[0072] Another type of measurement is an electrochemical impedance spectroscopy measurement, which may be carried out by loading the battery with a sinusoidal voltage the frequency of which varies within a given frequency range. FIG. 5 illustrates one example of a Nyquist plot of the measured electrochemical impedance, with the real part of the impedance on the abscissa and the opposite of the imaginary part of the impedance on the ordinate. Each point of the curve corresponds to a different excitation frequency.

[0073] All of the points on the curve may be selected or, advantageously, just certain characteristic points, for example the resistances Re_1 k, Re_10 measured at the frequencies 1 kHz and 10 Hz give a measurement of the internal resistance of the component and a measurement of the capacitive increase, respectively. These two measurement points are particularly relevant for assessing the correct operation of a microbattery.

[0074] In the case of microbatteries manufactured on silicon wafers, another collected datum is the position (x,y) of each component on the wafer.

[0075] Another collected datum relates to the manufacturing step at the end of which the test is carried out. In particular, the measurements may be carried out as soon as the three main elements of the battery, the positive electrode, the negative electrode and the contacts, have been manufactured. The measurements may notably be carried out before the manufacture of the encapsulation layers that protect the component from air and humidity.

[0076] All of the collected measurements are compiled in a training database of the type described in FIG. 6, in which each row corresponds to a tested component and each column corresponds to one of the measured parameters.

[0077] Once all of the measurements have been carried out, the method moves to step 111, which consists in carrying out complete cycling of the batteries, comprising a complete charge and/or a complete discharge.

[0078] During the cycling, a charge curve and/or a discharge curve is measured. The charge curve is for example a charge capacity or current curve. The discharge curve is for example a discharge capacity or voltage curve.

[0079] For each battery, the one or more measured curves are compared with reference curves corresponding to a functional state of the battery, and this is used to deduce a classification of each battery into the functional category (if the measurements correspond to the reference curves) or into the defective category (if the measurements do not correspond to the reference curves).

[0080] This step corresponds to a step of annotating or labelling the training data with a view to carrying out learning on the classification model. It may be carried out manually or automatically.

[0081] FIG. 7 shows various curves of charge or discharge profiles of a microbattery for characterizing functional components and defective components.

[0082] FIG. 701 shows one example of a charge profile curve in the form of a curve of the variation of the strength of a charge current as a function of charge capacity. The current evolution profile decreases as a function of charge capacity.

[0083] FIG. 702 shows one example of a discharge profile curve in the form of a curve of the variation of the voltage across the terminals of the battery as a function of discharge capacity. The voltage evolution profile decreases as a function of discharge capacity.

[0084] FIGS. 711, 712, 713, 714 show various charge profiles measured for defective chips. FIGS. 721, 722, 723, 724 show the corresponding discharge profile curves. Curves 711, 721 correspond to an open-circuit defect for which no current is measured during charging and the voltage across the terminals of the battery is zero during discharging. Curves 712, 722 correspond to a defect related to an excessively high current during charging. The charge curve 712 is not compliant as the measured current is too high, while the discharge curve is partially compliant. Curves 713, 723 correspond to a low-efficiency defect for which the charging of the battery is compliant with what is expected but the discharging is non-existent. Lastly, curves 714, 724 correspond to a metal layer breakage defect for which the current experiences a sudden interruption during charging and no voltage is able to be measured during discharging.

[0085] Other examples of charge and/or discharge profiles corresponding to defective chips may exist without departing from the scope of the invention.

[0086] By comparing at least the charge curve and/or at least the discharge curve with the reference curves 701, 702, it is deduced therefrom whether the battery is functional or defective. The curves may be compared visually or automatically, for example by comparing some or all of the points of the curves with one another.

[0087] This gives, at the end of step 111, a training database containing, for each tested microbattery, all of the associated measurements carried out in step 110 and the category “functional” or “defective” determined in step 111.

[0088] Based on these training data, supervised training 112 is then carried out on a classification model, the purpose of which is to determine the category of a microbattery based only on the measurements carried out in step 110, and to do so without carrying out complete cycling of the battery.

[0089] The supervised training 112 may be carried out using any unsupervised training engine or algorithm. A description is given of one exemplary embodiment of this training implemented by way of a random forest algorithm. Other machine learning algorithms may be envisaged, for example deep neural networks.

[0090] The data obtained at the end of step 111 are separated into a training dataset and a validation dataset. The machine learning model implemented via a random forest decision tree algorithm is trained via the training data. The training consists in iteratively adjusting the parameters of the model in order to be able to predict the functionality category of the microbatteries as precisely as possible.

[0091] The algorithm consists in constructing multiple decision trees by iteratively adding successive nodes. Each node is associated with a decision rule to be tested against input data. One example of a decision rule consists in comparing a value of an input datum with a threshold.

[0092] During the training, each decision tree is constructed from a subset of the initial dataset that is of the same size as the initial dataset but that comprises randomly drawn and replaced samples. In other words, one and the same sample (corresponding to a battery) may be represented multiple times in the subset used for training.

[0093] For each decision rule, a cost is computed to evaluate the effectiveness of the rule for segmenting the data associated with the node into the targeted categories.

[0094] The cost of a node is for example computed by way of the Gini criterion:

[0095] GiniIndex=1−Σ.sub.jp.sub.j.sup.2, where p.sub.j is the probability of a battery associated with a node belonging to the category j, j being equal to 0 or 1 in the present case where only two categories are targeted.

[0096] FIG. 8 illustrates one example of constructing successive nodes. The node 800 is associated with a rule for comparing the first datum X[0] (here the value of the open-circuit voltage) with a threshold value of −0.001 V.

[0097] For each node, the number of “unique” samples that satisfy the associated rule and the total number of samples (including the multiple samples) that belong to the two categories (defective or functional) is identified, this number being represented by the field “value=[x,y]”, where x is the number of samples belonging to the detective category and y is the number of samples belonging to the functional category.

[0098] Picking up on the example of FIG. 8, among the 73 samples associated with the node 800, 7 do not satisfy the inequality of the node 800 and are associated with the node 801, while 66 satisfy the inequality of the node 800 and are associated with the node 802. Among the 73 samples of the node 800, only 44 are “unique” samples, the others being copies of these “unique” samples.

[0099] Among the 7 batteries that do not satisfy the inequality X[0]<=−0.001 (node 801), 5 batteries belong to the defective category and 2 batteries belong to the functional category. The population of each category is roughly equivalent and the computed Gini index is 0.408, which is relatively high knowing that the maximum value of this index is 0.5.

[0100] By contrast, among the 66 batteries that satisfy the inequality X[0]<=−0.001 (node 802), 65 belong to the defective category and just one belongs to the functional category, thereby leading to a low Gini index, equal to 0.03.

[0101] The cost of a rule is given by the following relationship:

[00001]$J\left(k\right)=\frac{m_{\mathrm{left}}}{m}{G}_{\mathrm{left}}+\frac{m_{\mathrm{right}}}{m}{G}_{\mathrm{right}}$

[0102] where J(k) is the cost of the node of rank k, m is the total population associated with the node, m.sub.left, G.sub.left, m.sub.right and G.sub.right are respectively the populations and Gini indices of the left and right sub-nodes created by the decision rule.

[0103] In the example above, the cost of the decision rule associated with the node 800 is therefore around 0.067. When training the model, for each rank, a large number of inequalities are tested randomly on all of the input data, and the one having the lowest cost is retained.

[0104] The construction of a new node of rank k+1 then continues restarting from the node, in this example the node 802, whose Gini index is lowest. Multiple criteria may be envisaged for continuing the construction of the following nodes: [0105] The maximum depth, that is to say the maximum number of levels that a tree is able to have, [0106] The minimum number of samples from which a node is able to be divided into sub-nodes, and [0107] The minimum number of samples in each sub-node after a division. For example, if this number is taken to be equal to 5, a division will be considered only if it results in the creation of 2 sub-nodes each having at least 5 samples.

[0108] FIG. 9 shows a partial example of a tree that is obtained.

[0109] The dashed nodes continue via other nodes that are not shown for the sake of clarity (the whole tree being too large). The final prediction of the tree is given by comparing the values measured for each battery with the inequalities of each node. Running through the tree from node to node arrives at a leaf node. Knowing the distribution of the categories for this leaf for the learning data, it is then possible to estimate a probability of belonging to one or the other of the categories, and the most probable category is assigned to the battery.

[0110] In other words, the category of belonging is the one for which the leaf node has the greatest number of samples during learning.

[0111] The random forest algorithm generates multiple trees of the type of FIG. 9 (of the order of one hundred for example) that are trained on all or some of the training data, so as to reduce the overfitting phenomenon. The final prediction of the model for a given battery is given by a majority vote expressed over all of the trees.

[0112] Optionally, the trained classification model may be evaluated on a validation dataset by computing a prediction score on the training data and a prediction score on the validation data in order to check whether the difference between the two scores is not too large. If this difference is too large, new learning may be carried out by modifying the parameters of the algorithm.

[0113] Once the classification model has been trained and validated, it is used directly in the use phase in step 114 to predict whether a new manufactured battery is functional or defective based only on the measurements 110 carried out on the battery without carrying out complete cycling.

[0114] FIG. 10 shows one example of predictions obtained on the scale of a silicon wafer on which microbatteries are manufactured. Functional batteries are identified by the category ‘1’, and defective batteries are identified by the category ‘0’.

[0115] FIG. 11 shows a diagram of a test system configured to implement the invention. To this end, FIG. 11 shows a battery Dn having an anode and a cathode, a programmable current generator An and a programmable voltmeter Vn that are connected to the terminals of the battery Dn in order to carry out the measurements needed in step 110 of the method according to the invention. The test system comprises a controller Cn in the form of a processing unit that receives the measurements carried out on the battery and implements the learning and the use of the classification model to predict the functional or defective state of the battery Dn. The controller Cn comprises at least one processor Pn and a memory Mn.

[0116] The invention may be implemented as a computer program comprising instructions for the execution thereof. The computer program may be recorded on a processor-readable recording medium.

[0117] The reference to a computer program that, when it is executed, performs any one of the functions described above is not limited to an application program running on a single host computer. On the contrary, the terms computer program and software are used here in a general sense to refer to any type of computer code (for example application software, firmware, microcode, or any other form of computer instruction) that may be used to program one or more processors to implement aspects of the techniques described here. The computing means or resources may notably be distributed (“Cloud computing”), possibly using peer-to-peer technologies. The software code may be executed on any suitable processor (for example a microprocessor) or processor core or a set of processors, be these provided in a single computing device or distributed among multiple computing devices (for example as possibly accessible in the environment of the device). The executable code of each program allowing the programmable device to implement the processes according to the invention may be stored for example in the hard drive or in read-only memory. Generally speaking, the one or more programs will be able to be loaded into one of the storage means of the device before being executed. The central processing unit is able to command and direct the execution of the software code portions or instructions of the one or more programs according to the invention, which instructions are stored in the hard drive or in the read-only memory or else in the other abovementioned storage elements.

[0118] The invention may be implemented on a computing device based for example on an embedded processor. The processor may be a generic processor, a specific processor, an application-specific integrated circuit (ASIC) or a field-programmable gate array (FPGA). The computing device may use one or more dedicated electronic circuits or a general-purpose circuit. The technique of the invention may be implemented on a reprogrammable computing machine (a processor or a microcontroller for example) executing a program comprising a sequence of instructions, or on a dedicated computing machine (for example a set of logic gates such as an FPGA or an ASIC, or any other hardware module).