Method for evaluating the capacity of electric vehicle battery cells using a statistical machine learning model

Abstract

The invention relates to a method for evaluating the capacity of electric vehicle or hybrid battery cells (1) during their manufacture. This method uses a statistical model (9) of the cells, trained by training cells, to perform an indirect evaluation by means of measurements of characteristics, particularly physical and electrical, of the cells during their manufacture. The lengthy and costly direct measurement of capacity, which involves fully charging and then discharging the cell, becomes unnecessary. The model is machine learning-based (13), capable of being amended to maintain its accuracy even if the cell characteristics drift. (Fig. 4)

Description

Title of the invention: Method for evaluating the capacity of electric vehicle battery cells using a machine learning statistical model. Technical field of the invention

[0001] The invention relates to rechargeable electric batteries, including those with high electrical capacity such as several hundred ampere-hours, for example.

[0002] It relates more particularly to a method for evaluating the discharge capacity of cells in such electric batteries used to power electric or hybrid vehicles, in which a statistical machine learning model is used to predict this capacity without having to measure it directly. Prior art

[0003] The manufacture of electric battery cells intended to power vehicles (often referred to as "cells" in the remainder of this text) includes an electrical processing phase, followed by a qualification phase, called "grading," designed to verify that a required electrical discharge capacity for the cell has been achieved. Finally, after rejecting non-conforming cells, the qualified cells are brought to a final state of charge before being removed from the factory.

[0004] The electrical treatment consists of a succession of preliminary applications of electric currents to the cells to induce a chemical transformation of the electrodes, in particular by forming a passivation layer, capable of enabling them to store electrical energy at the desired capacity.

[0005] In traditional processes, the qualification phase mainly consists of fully charging the cells and then completely discharging them. The discharge capacity of the cells is measured based on the time required for this complete discharge, known as "grading discharge capacity" (GDC), from the maximum charge state. Finally, a partial recharge must be performed to reach the final charge state mentioned above.

[0006] After the long and energy-intensive electrical treatment, this final group of steps is itself expensive and relatively long.

[0007] In order to shorten and simplify the qualification phase, the applicant has designed an improved method whereby a machine-learning statistical model is used to predict the discharge capacity of cells, without having to physically measure it by a final complete discharge. This model is built in a preliminary training or conditioning phase, exploiting the characteristics of training cells that are as similar as possible to the cells whose capacity is to be predicted.

[0008] While the statistical model can, in principle, predict cell capacity with sufficient accuracy, its reliability may be called into question if deviations in the characteristics of the cells being evaluated compared to the training cells are observed. Such deviations are inevitable and are likely to be observed in practice, even in small production runs. These deviations result in particular from changes in the components and raw materials used, alterations in the manufacturing processes and external conditions, or deliberate improvements to the cell manufacturing process. They may also result from obtaining more data. Description of the invention

[0009] The invention was designed to overcome this inadequacy of the statistical model, and to allow its confident use even on long series of cells to be evaluated, despite the inevitable drifts in their characteristics.

[0010] In general terms, the invention relates to a method for evaluating the capacity of electric battery cells for electric or hybrid vehicles, characterized by the following steps:

[0011] a) measurements of parameters characterizing cells to be evaluated at least during a phase of a manufacture of said cells to be evaluated, during which an electrical treatment to form their capacity is applied to them;

[0012] b) introduction of input variables of the cells to be evaluated in a statistical model, the input variables being taken from parameter measurements;

[0013] c) prediction of the cells' ability to be evaluated by the statistical model according to the input variables,

[0014] in which the statistical model is built through training phases, using a database comprising, for manufactured training cells, input variables of the training cells homologous to the input variables of the cells to be evaluated, and measures of the capacity of the training cells,

[0015] and the process further comprises the following steps:

[0016] d) estimation of drifts of the input variables of the cells to be evaluated with respect to the input variables of the training cells;

[0017] e) sampling of cells to be evaluated whose drifts in input variables exceed a threshold;

[0018] f) measurement of the capacities of the cells to be evaluated taken, when the manufacture of said cells is finished;

[0019] g) entry into the database of input variables and measured capacities of the cells to be evaluated taken, obtaining a completed database;

[0020] h) amendment of the statistical model by a supplementary training phase using the completed database.

[0021] In this text, "homologous" means "of a similar nature, but with generally different values." One property of the statistical model is that it is easy to build automatically, through simple processing of the database. It can therefore be easily amended at will, as soon as the database is completed. The process constantly monitors for the occurrence of excessive drifts and triggers additional model training as soon as such excessive drifts appear.

[0022] In practice, it is advantageous for the largest number of input variables, or even all of the input variables, to participate in the detection of excessive drift, and therefore for the drifts to be estimated on each of the input variables.

[0023] The additional training can be ordered either based on the drift of each of the input variables, or based on one or more global drifts calculated on one or more groups or on all the input variables, or based on a combination of these criteria. A specific threshold is then assigned to each input variable or each group of these variables as appropriate.

[0024] In important embodiments, the detection of excessive drift is not performed individually on each cell. Instead, the cells to be evaluated are grouped into batches, the input variables of the cells to be evaluated in the batch are arranged into statistical distributions, and the drifts are estimated by comparing said statistical distributions to homologous statistical distributions obtained on the input variables of the training cells.

[0025] Among other possibilities, said batches of cells to be evaluated may be limited by a sliding time window during which the cells to be evaluated from the same batch reach the same manufacturing state.

[0026] Among other possible criteria, Kolmogorov-Smirnov tests are particularly considered for comparing statistical distributions.

[0027] It should be emphasized that the evolution of the model can be done in masked time with the main evaluation process, that is to say that steps (d) to (h) are carried out at the same time as steps (a) to (c), without interrupting the manufacture of the cells.

[0028] Another method of amending the model, usable with the previous one, further comprises the following steps:

[0029] i) estimation of drifts in the capacities of the cells to be evaluated compared to the capacities of the training cells;

[0030] j) sampling of second cells to be evaluated whose capacity drifts exceed a second threshold;

[0031] k) measurement of the capacities of the second cells to be evaluated taken, when manufacturing is finished;

[0032] 1) entry into the database of input variables and measured capacities of second cells to be evaluated taken, obtaining a completed database;

[0033] m) amendment of the statistical model by a supplementary training phase using the completed database.

[0034] A third method of modification, usable with the previous ones, is also envisaged. It differs from them in that it does not consist of ordering enrichments of the database through new calibrations, nor additional model training. Instead, it comprises a step of estimating the error of the capacity prediction, for cells to be evaluated whose capacity has been measured, and a step of adjusting said threshold and / or said second threshold according to said error.

[0035] The main advantage of such adjustments is to limit untimely alerts by which additional training would be requested when the model is actually satisfactory, giving good predictions of capabilities despite the existence of drifts on the characteristics of the cells.

[0036] The adjustment criterion advantageously includes three situations: a lowering of the threshold or thresholds if the error is judged excessive, an raising if the error is judged small, no adjustment being made if the error is moderate, between excessive and small. Brief description of the figures

[0037] The present invention will be better understood upon reading the description of non-limiting examples of embodiments, with reference to the accompanying figures, which illustrate: • [Fig.1]: a diagram of a battery subjected to electrical treatment and measurement; • [Fig.2]: the evolution of certain characteristic parameters of the process in accordance with the prior art, in the form of time functions; • [Fig.3]: the modifications to these parameters that the invention implies; • [Fig.4]: the complete installation which carries out the process; • [Fig.5]: a characteristic drift in a series of cell manufacturing; • [Fig.6]: an organizational chart of the capacity assessment process; • [Fig. 7]: A flowchart of a first method of amending the model statistics; • [Fig-8]: an organizational chart of a second method of amending the statistical model; • [Fig. 9]: A flowchart of a third method of amending the model statistics; • [Fig. 10]: an illustration of statistical distribution. Detailed description of at least one embodiment

[0038] For clarity, identical elements are identified by identical reference numerals from one figure to another. The examples given are purely illustrative; variations are within the grasp of a person skilled in the art without departing from the invention.

[0039] Figure 1 schematically illustrates the installation required for cell production during the electrical processing and qualification phases. A cell 1 has its terminals 2 connected to an electrical circuit 3 equipped, in particular, with a DC generator 4 capable of delivering current in either direction, and with means 5 for measuring the electrical parameters in the circuit 3, notably the current intensity and the voltage between the terminals 2. Other measuring means 6 are fitted to the cell 1. These measuring means 6 may relate, in particular, to the temperature of the cell 1, its internal pressure, its dry weight, or to physical or chemical properties of the electrolyte, such as its weight. The measurements may also be taken by different means, before or during manufacturing.

[0040] Figure 2 illustrates the execution of a concrete process of a known kind, often called a "recipe" in the prior art. It is a time-domain current profile, represented here in the form of three diagrams illustrating, from top to bottom of the figure, three of these parameters, namely the voltage V applied to terminals 2 by circuit 3, the electrical charge Q that can be applied to or withdrawn from cell 1 during each charging or discharging step (knowing that Q = JI during the step, where I is the intensity of the applied current), and then the intensity I itself; a fourth diagram is that of the temperature °T of cell 1.The process lasts several days, denoted J1 to J5, and essentially consists of applying four constant-intensity currents and one variable-intensity current at constant voltage during as many training stages F2, F3, F4, and F' on day J2; then three final constant-intensity currents and another constant-voltage current during the last training stages F' and F5; and finally two qualification stages G6 and G7 on the last day J5. Many other methods are known in the prior art, each of which could be improved by the invention. The current applications can generally be consecutive or, conversely, separated by rest periods, with currents flowing in either direction. with potentially all different intensities and durations. Non-uniform currents can also be considered.

[0041] The first qualification step G6, carried out under negative direction current, corresponds to the complete discharge of cell 1, and the second qualification step G7 to the final partial recharge step, up to the imposed final state, at the charge rate Co. These steps are carried out to completion in traditional processes.

[0042] The invention, however, makes it possible to interrupt the process at the said stopping instant ti, where the cell 1 already reaches this imposed state during the first qualification step G6, which is therefore abbreviated, while the second qualification step G7 is omitted entirely.

[0043] Figure 3 illustrates the qualification phase when the process according to the invention is used, with diagrams of the applied electrical charges Q and the current intensity I. It consists of a single step, denoted G'6, in which the cells 1 are directly brought to the final charge state Co after the electrical treatment phase, i.e., without going through complete discharge or final recharging.

[0044] The overall installation is illustrated in [Fig. 4]. The cells 1 undergo the preparation process on a production line 8 up to time tb. A computer server 9 contains a statistical model that receives the values ​​of the input variables from measurements taken by means 5 and 6 on the production line 8 up to time ti, or by other means, elsewhere or earlier. The model then predicts the capacity of each of the cells 1. When a batch 10 of the cells 1 has been processed, their capacity predictions and possibly other characteristics are provided to a first control station 11. The cells 1 deemed non-conforming are identified and withdrawn from the market.The computer server 9 also communicates with a second control station 12, for example to obtain an evaluation of the model, which is periodically retrained using new calibration cells 1 in the manner to be developed, particularly if its predictions become inaccurate. For this, a training module 13 is used, already used for the preliminary calibration of the model.

[0045] Thus, the physical measurement of capacity, typical of traditional processes, is replaced by a prediction made by the statistical model using other measurements, easy and inexpensive to carry out.

[0046] The general characteristics of the process performed by the statistical model from the measurements made on cells 1, as well as those of the statistical model itself, will now be described.

[0047] Measurements on cells 1 can be taken periodically or not between the initial time t0 of the process and the end of the processing phase, and preferably until the stopping time tb. However, while these measurements can be very numerous, i.e., up to several tens of thousands for a single cell 1, it is necessary for the model to use only a limited number of input variables due to its limited computing power and the expected loss of learning performance if the variables are too numerous. Furthermore, most measurements are either of little use or redundant in many cases. A selection process is carried out to choose as input variables only those best suited to predicting capacity, starting with those most strongly correlated with it.The desire to avoid using redundant measures, however, leads to the elimination of variables that are too highly correlated with others.

[0048] The choice of input variables for the model, and of the measurements used to form these input variables, can be made by artificial intelligence during the initial calibration phase of the model, which is performed on cells 1 whose capacity has also been measured according to the ordinary process. The measurements most suitable for predicting capacity may differ for each model of cell 1 and each recipe. However, it has been noted that measurements more particularly likely to serve as input variables for the model include, at least for some, steps of the electrical processing which are called Fi to F5 in the example in [Fig. 2]:

[0049] -electrical charges and discharges Q: the total value applied during the step, the average value during the step, or the time derivative;

[0050] -voltage V: the maximum during the step, the average value during the step, or the time derivative;

[0051] -internal pressure: the average during the step;

[0052] -temperature °T: the maximum or average during the step;

[0053] -current I: the average value or the time derivative during the step;

[0054] -the dry weight of the cell;

[0055] -the weight, or other physical or chemical characteristics, of the electrolyte.

[0056] This list is not exhaustive. It includes both direct measurements and indirect measurements obtained by processing the direct measurements, for example through integration or numerical differentiation. The input variables may be associated with measurements taken at specific times, time intervals, or variable times (for extrema, for example). It should also be noted that while many of these measurements are taken during the electrical processing, others are taken, or may be taken, before this processing, such as weighings.

[0057] The variables resulting from the selection of measures can number from approximately 10 to 40 in plausible implementations. These variables are provided at the input of the model, which directly determines the capacity of the cell 1 considered, the duration of the single qualification step, and the stopping time ti of the manufacturing process.

[0058] It has been found that such a small number of input variables of the model is sufficient to deduce the capacity of a cell with a high accuracy of approximately ±0.2%, including for non-conforming cells, with capacities below the standards called "scraps".

[0059] The statistical model can be of various known kinds: decision tree, deep learning, or regression in particular.

[0060] If a decision tree is chosen, the algorithm will consist of successively comparing the input variables, or only some of them depending on the branch of the tree, to respective thresholds, so as to evaluate the capacity of the corresponding cell with increasing precision. The thresholds are estimated during the calibration of the model, as are the variables used in each branch. Typically, the variables used first will be the most discriminating, that is, those most strongly correlated with capacity.

[0061] Among the many variants of decision trees, the random forest seems very interesting here, the decision being made on several trees at the same time.

[0062] Deep learning can consist of using a neural network, for example, a multilayer perceptron. The input variables of each layer are transformed according to the model y = o(Sx+b), where x is the input vector, y the output vector of the layer, o a discrimination function, for example a sigmoid, S a weighting matrix, and b a bias matrix. The input variables of the model constitute the input vector of the first layer, and the capability evaluated by the model is found in the output vector of the last layer.

[0063] Regression algorithms are based in particular on classical linear regression or the use of support vectors. They consist of assigning weighting coefficients to the input variables, enabling the direct calculation of capacity.

[0064] The model can also be hybrid or composite, that is to say, made up of several algorithms of different kinds working simultaneously, or of which on the contrary only one is chosen as the best to evaluate the capacity, according to the results of the preliminary calibration, according to the characteristics of the battery.

[0065] In all cases, the statistical model can be built during a preliminary training phase in an automated manner, that is to say by an artificial intelligence program which determines, at the same time as the input variables useful for the input of the statistical model, the constitution and parameterization of the statistical model itself, and for example the tree rules, the numerical thresholds that these rules use, the number of layers of the neural network, the coefficients weighting in S-weighting matrices or regression algorithms, etc. This training phase uses training cells, in principle similar to the cells to be evaluated later; the capacity of the training cells is known by being measured directly using traditional methods, by a complete discharge. The model is therefore built empirically, through successive optimizations and adjustments, so that the input variables of the training cells at the model's input yield the capacities measured at its output.

[0066] The high accuracy that can be expected with such statistical models depends, however, on the quality of the sampling of the training cells. If the input variables of the cells to be evaluated differ from those of the training cells, the model's prediction of the capacity of the cells to be evaluated is no longer guaranteed. However, the cells constantly undergo significant drifts in their input variables and capacities due to changes in the actual manufacturing conditions, the raw materials used, and many other poorly understood parameters. These drifts are unpredictable and can appear either slowly over large series of successive cells or abruptly from one cell to the next.

[0067] Figure 5 illustrates these drifts. Cells 1, successively manufactured (at different times t, on the x-axis), obtain predictions of their capacity (GDC, on the y-axis) using the process of the invention. These predictions correspond to points 21. The line 20 connecting them is broken with a highly irregular shape, illustrating the erratic and unpredictable nature of the drifts. The upper and lower envelopes 22 and 23 represent the confidence limits of the predictions. The actual capacity values ​​of these cells 1 were also physically measured: the results are indicated by the vertical bars 24, the central point of which indicates the measurement value, and the height of which expresses the measurement uncertainty. The figure is given for illustrative purposes only, as the drifts affect not only the capacity value but also all manufacturing parameters and all input variables of the model.

[0068] The method of the invention, by which the statistical model is made scalable by machine learning by which it is amended according to the observed drifts, will now be explained by means of the flowcharts in figures 6 to 9.

[0069] According to [Fig. 6], for each of the cells 1 to be evaluated, it involves regular measurements of the cell 1 parameters during its manufacture (step E1). When manufacturing is complete, and the qualification phase can begin (step E2), the values ​​of the input variables of cell 1 are determined (E3) by selecting the measurements or by digitally processing these measurements according to the possibilities indicated above. These values ​​of the input variables are then provided to the statistical model (E4), which makes a prediction of the capacity of cell 1 (E5). Cell 1 is then accepted or rejected based on the prediction. capacity (E7), unless a capacity measure is decided (E6); this situation will be detailed later.

[0070] In parallel with this main part of the evaluation process, the learning of the statistical model is ensured by means of several amendment modes, which can be present simultaneously.

[0071] The first is illustrated in [Fig. 7]. It is based on the input variables and begins with a branch at step E3. The input variables of the cell 1 processed up to this step E3 are incorporated into a batch that also includes the input variables of other cells 1 (step E8), which may be cells 1 created during the same sliding time window, for example. When the batch of input variables is complete, it is arranged into statistical distributions for each of the input variables (E9). A statistical distribution, as the pictogram in [Fig. 10] reminds us, is composed of classes 27 such as those shown in [Fig. 10], each of which expresses that the input variable has a probability p of having the value V.

[0072] When the statistical distributions of the input variables for the set of cells 1 are established, homologous distributions, that is, distributions based on the same input variables but for training cells previously recorded in a model database (E10), are compared (E11) to said distributions for cells 1, by means of a numerical test. Several exist in the field of statistics, of which the Kolmogorov-Smimov test is particularly considered here. It consists of calculating, for two statistical distributions each comprising n probability classes pl1 and p2, and for each of the classes i between 1 and n, the total difference E(i) between the two distributions from one of the endpoints: E(i) = Sj [(pl(j) _ p2(j)], where j varies from 1 to i.

[0073] Regardless of the test used, it is then determined whether excessive drift of the input variables is present between the cells to be evaluated in the batch and the training cells (E12). The test results are compared to a threshold. In the usual case where there are several input variables, they are advantageously, but not necessarily, all taken into account to detect excessive drift. When several input variables are taken into account, the drifts of each can be summed, after being weighted, before making the comparison to a single threshold. In the case of a Kolmogorov-Smirnov test giving several deviation results for each of the distributions, the largest deviation E(i) of each pair of distributions compared in the test can be chosen to be included in the comparison with the threshold.

[0074] If the threshold is not exceeded, no action is taken and the process started in step E8 is finished for this cell 1 (El3). If the threshold is exceeded, a A set of cells to be evaluated, which become supplementary training cells (El4), are added to the model database to complete it. This set of cells to be evaluated may correspond to the set from step E8 that formed the statistical distributions of the input variables, or it may be different. It is important that this set contains at least some of the cells to be evaluated for which excessive drifts were observed. The capacity of the supplementary training cells must also be measured directly, as with the previous training cells.

[0075] The final step E15 of this part of the process is a further training phase of the statistical model using the additional training cells, in order to amend the model by supplementing it in a region of the input variable values ​​where it had not yet been tested. The amended statistical model can then be used immediately on the cells 1 currently being processed and awaiting the capacity prediction step E5.

[0076] A second method of amending the statistical model, some steps of which are analogous to those of the first method, begins with a branch at step E5; it is described by means of [Fig.8].

[0077] The first step (E16) consists of incorporating the predicted value of the capacity of cell 1 into a batch that also includes the predicted values ​​for other cells 1. When the batch is large enough to be considered complete, a statistical distribution of the predicted capacity values ​​is established (E17), a corresponding distribution for the capacities of the training cells is extracted from the database (E18), and the two distributions are compared (E19). The Kolmogorov-Smimov test can also be used for this purpose. If the difference between the two distributions does not exceed a threshold (E20), this part of the process ends (E21).If it exceeds this limit, the process concludes that there is excessive drift in the predicted capacity values, organizes a batch of cell 1s to form additional training cells (E22), which are incorporated into the database and used here as well for a new phase of training the statistical model (E23). In a final step (E24), the statistical model is amended and immediately usable for the cell 1s currently being manufactured. This amendment method is very similar to the previous one, except that the drifts are observed in the capacities rather than in the input variables. The remarks made about the first method remain valid here, particularly regarding the composition of the batches and the choice of test.

[0078] A third method of amending the statistical model exists, and it is detailed by means of [Fig. 9]. It begins with a branching at step E6, if a measurement of the capacity of cell 1 is decided upon, for example at predetermined intervals. When the measurement is made, cell 1 is accepted or rejected like the others at Step E7, however, is based on the measurement of capacity rather than its prediction. In parallel, the process calculates a value for the difference between the measured and predicted capacity values ​​in step E24, i.e., a prediction error. It incorporates this error value into a batch of corresponding values, i.e., prediction-measurement errors calculated for other cells (E25), and determines whether the batch's statistical error is excessive, low, or moderate in the following step E26, once the batch is complete. This statistical error can be composed of both the mean and the root mean square of the error values ​​within the batch. The result is used to adjust the drift detection thresholds, which are applied in steps E12 and E20 described previously.If the statistical error is small, it is deduced that the statistical model is very satisfactory, requires little modification, and that the thresholds for detecting deviations can be raised to reduce the risks of imposing additional training that might be superfluous; conversely, they are lowered if the statistical error is excessive and suggests that the statistical model is not precise enough, and that additional training should be encouraged; finally, the thresholds remain unchanged if the statistical error has a moderate value, which is considered normal.

[0079] We have considered here processes using batches of cells 1. These batches can be constructed in different ways, composed of cells 1 manufactured for example during a sliding window of time or at predetermined intervals; the number of cells 1 in each batch is not critical, and it can be different for each batch.

[0080] In accordance with the above, the training module 13 comprises, according to [Fig.4]: an input variable drift controller 14, which therefore receives the input variables from the relevant cells 1; a capacity drift controller 15, which receives the predictions and capacity measurements from the cells 1 involved in the last two amendment processes; a drift detector 16, sensitive to the indications provided by the controllers 14 and 15; and the training program 17, controlled by the drift detector 16. The device also includes the database, which is not shown in the figure because it can be housed in completely separate hardware, or even a data cloud.

Claims

Demands

1. A method for evaluating the capacity of electric battery cells for electric or hybrid vehicles, characterized by the following steps: a) measurements (E1) of parameters (I, Q, V, °T) characterizing cells (1) to be evaluated at least during one phase of a manufacturing of said cells to be evaluated, during which an electrical treatment to form their capacity is applied to them; b) introduction (E4) of input variables of the cells to be evaluated into a statistical model (9), the input variables being taken from the parameter measurements;c) prediction (E5) of the capacity of the cells to be evaluated by the statistical model according to the input variables, in which the statistical model (9) is built by training phases, using a database comprising, for manufactured training cells, input variables of the training cells homologous to the input variables of the cells to be evaluated, and measurements of the capacity of said training cells, and the process further comprises the following steps: d) estimation (E8) of drifts of the input variables of the cells (1) to be evaluated with respect to the input variables of the training cells; e) sampling of cells to be evaluated whose input variable drifts exceed a threshold (E12); f) measurement of the capacities of the sampled cells to be evaluated, when the manufacture of said cells is finished;g) input (El4) into the database of input variables and measured capacities of the cells to be evaluated, obtaining a completed database; h) amendment (El5) of the statistical model (9) by a further training phase using the completed database.

2. Method for evaluating the capacity of electric battery cells for electric or hybrid vehicles according to claim 1, characterized in that the drifts are estimated on each of the input variables.

3. Method for evaluating the capacity of electric battery cells for electric or hybrid vehicles according to any one of claims 1 or 2, characterized in that the additional drive phase is controlled according to the drift of each of the input variables and / or according to one or more global drifts calculated on one or more groups or on all of the input variables.

4. Method for evaluating the capacity of electric battery cells for electric or hybrid vehicles according to any one of claims 1 to 3, characterized in that the cells to be evaluated are grouped into lots (10), the input variables of the cells to be evaluated in the lot are arranged into statistical distributions (27), and the drifts are estimated by comparing said statistical distributions to homologous statistical distributions obtained on the input variables of the drive cells.

5. A method for evaluating the capacity of electric battery cells for electric or hybrid vehicles according to claim 4, characterized in that the batches of cells to be evaluated are limited by a sliding time window during which the cells to be evaluated from the same batch reach the same manufacturing state.

6. Method for evaluating the capacity of electric battery cells for electric or hybrid vehicles according to any one of claims 4 or 5, characterized in that the statistical distributions are compared (Eli) according to Kolmogorov-Smirnov tests.

7. Method for evaluating the capacity of electric battery cells for electric or hybrid vehicles according to any one of claims 1 to 6, characterized in that steps (d) to (h) are carried out at the same time as steps (a) to (c).

8. A method for evaluating the capacity of electric battery cells for electric or hybrid vehicles according to any one of claims 1 to 7, characterized in that it further comprises the following steps: i) estimation (E16, E19) of capacity drifts of the cells (1) to be evaluated with respect to the capacities of the drive cells; j) sampling of second cells to be evaluated whose capacity drifts exceed a second threshold; k) measurement of the capacities of the second cells to be evaluated taken, when manufacturing is finished; 1) entry (E21) into the database of the input variables and the measured capacities of the second cells to be evaluated taken, obtaining a completed database; m) amendment (E22) of the statistical model by a supplementary training phase using the completed database.

9. Method for evaluating the capacity of electric battery cells for electric or hybrid vehicles according to any one of claims 1 to 8, characterized in that it comprises a step (E24) of estimating the error of the capacity prediction, for cells (1) to be evaluated whose capacity has been measured, and a step (E27) of adjusting said threshold and / or said second threshold as a function of said error.

10. Method for evaluating the capacity of electric battery cells for electric or hybrid vehicles according to claim 9, characterized in that the adjustment consists of a lowering if the error is judged to be excessive, an raising if the error is judged to be low, no adjustment being made if the error is moderate, between excessive and low.

Images