A computer processing method involving a neural network, device and associated system.

Abstract

A computer processing method comprising a repetition of the following first steps: Receiving a first set of values ​​for a driving parameter of a motor vehicle (100), in a first chronological order; Determining a first sample from the first set, the first sample comprising a plurality of first values ​​from the first set, in the first chronological order, each value in the plurality of first values ​​being preceded by previous first values ​​in the first set, and followed by subsequent first values ​​in the first set, the previous and subsequent first values ​​being excluded from the first sample; First supervised training of a first artificial neural network (111) comprising one input and one output, from the first sample received at the input, the first sample being labeled by the first set. Figure for the abstract: Figure 1

Description

Title of the invention: Computer processing method involving a neural network, associated device and system.

[0001] The invention relates to series of values ​​of a driving parameter of a motor vehicle collected during a journey of the motor vehicle. There is a need to reduce the volume of data that these series can constitute, in order to exchange them, store them, or more generally use them.

[0002] For this purpose, the invention relates to a computer processing method comprising a repetition of the following first steps, for example a number of times: • Reception of a first series of values ​​of a driving parameter of a motor vehicle, in a first chronological order, then • Determining a first sample of the first series of values, the first sample comprising a plurality of first values ​​(e.g., at least 5 first values) of the first series of values, in the first chronological order, each value of the plurality of first values ​​being preceded by preceding first values ​​in the first series of values, in the first chronological order (e.g., contiguous values ​​in the first series, e.g., at least 9 contiguous values ​​in the first series), and followed by following first values ​​in the first series of values, in the first chronological order (e.g., contiguous values ​​in the first series, e.g., at least 9 contiguous values ​​in the first series), the preceding first values ​​and the following first values ​​being excluded from the first sample (i.e., not in the first sample, not belonging to the first sample,or are outside the first sample), then, • First supervised training (in other words: second learning) of a first artificial neural network (in other words: of a first electronic neural network) comprising an input and an output, from the first sample, (received) at the input (, by the first artificial neural network), the first sample being labeled by the first series of values, the first artificial neural network comprising synaptic weights, the first supervised training modifying the synaptic weights.

[0003] Thus, according to the invention, a first artificial neural network is trained which will be able to produce (in other words: to predict) a series of values ​​similar to the first series from the first sample, which can make it possible to memorize or transmit only the first sample instead of the first series.

[0004]

[0005]

[0006]

[0007]

[0008] The number of times depends on the architecture of the first artificial neural network, the data at the input of the artificial neural network during training and the expected quality of the predictions made by the first artificial neural network. It can be between 2 and 1000, for example. According to one embodiment, the method further comprises the following second steps, possibly repeated, implemented in the motor vehicle, after the repetition of the first steps: • First transmission of synaptic weights to a computer device, then • Receipt of a second series of values ​​of the motor vehicle's driving parameter, in a second chronological order, then • Determining a second sample of the second series of values, the second sample comprising a plurality of second values ​​(e.g., at least 10 second values) of the second series of values, in the second chronological order, each value of the plurality of second values ​​being preceded by previous second values ​​in the second series of values, in the second chronological order (e.g., by contiguous values ​​in the second series, e.g., by at least 9 contiguous values ​​in the second series), and followed by next second values ​​in the second series of values, in the second chronological order (e.g., by contiguous values ​​in the second series, e.g., by at least 9 contiguous values ​​in the second series), the previous second values ​​and the next second values ​​not belonging to the second sample (in other words: are not in the second sample,or are excluded from the second sample, or are outside the second sample), then, • Transmission of the second sample to the computing device (for example, a server or computer). This helps reduce the volume of information sent to the computing device. Alternatively, the first neural network can be used to store the second set of values ​​using reduced memory space (by storing only the first sample). According to one embodiment, during the reception of the first series, the first series is received, for example, with a period, for example between 100 hertz and 1 hertz (in a variant in a non-periodic manner), from a sensor or several sensors of the motor vehicle (in a variant, from a radiofrequency communication means, for example from a geographic location module satellite, or according to another variant, from an image processing module of a camera of the motor vehicle), for example when the motor vehicle is moving.

[0009] The first chronological order is for example the order in which the rolling parameter was measured by the sensor or said several sensors, or the order in which the values ​​of the first series are received, during the step of receiving the first series. Other variants are possible, of course.

[0010] According to one embodiment, during the reception of the second series, the second series is received, for example with a period, for example between 100 hertz and 1 hertz (in a variant in a non-periodic manner), from a sensor or several sensors of the motor vehicle (in a variant, from a radiofrequency communication means, for example from a satellite geographic location module, or according to another variant, from an image processing module of a camera of the motor vehicle), for example when the motor vehicle is moving.

[0011] The second chronological order is for example the order in which the rolling parameter was measured by the sensor or said several sensors, or the order in which the values ​​of the second series are received. Other variants are possible.

[0012] The plurality of first values ​​may be distributed homogeneously in the first series. For example, the step of determining the first sample may comprise a step of selecting the first values, the first values ​​being one value out of 10 of the first series, in chronological order. Other distributions, for example heterogeneous, of the first values ​​in the first series, and other methods of selecting the first values ​​may of course be envisaged. The same applies to the second values ​​and the second sample.

[0013] According to one embodiment, the method further comprises the following third steps, implemented, in the motor vehicle, after the step of determining the second sample of the second series of values ​​and before the transmission of the second sample: • First production (in other words: prediction), by the first artificial neural network, of a third series, at the output of the first artificial neural network, from the second sample, (received) at the input of the first artificial neural network, then • Checking whether a distance (e.g., Euclidean, alternatively, a so-called “Manhattan” distance, a so-called “Minkowski” distance or a so-called “Chebyshev” distance) between the second series and the third series is greater than a (predetermined) threshold, • If said distance is greater than the threshold (and these steps can be repeated as long as the distance is not less than the threshold): • Second training (in other words: second learning), supervised of the first artificial neural network from the second sample, (received) at the input of the first artificial neural network, labeled by the second series, the second training modifying the synaptic weights of the first artificial neural network then, • Second transmission of synaptic weights to the computer device.

[0014] The implementation of the first supervised training and the second supervised training is well known to those skilled in the art. It can be carried out, for example, by a gradient backpropagation method (alternatively, for example, by a calculation of the second derivative).

[0015] For example, the first transmission of the synaptic weights, and / or the transmission of the second sample, and / or the second transmission of the synaptic weights and / or the reception of the synaptic weights hereinafter and / or the reception of the second sample hereinafter, may be implemented via radiofrequency communication, for example in a mobile telephone network, and / or for example, via a mobile telecommunications module of the motor vehicle (of the computing device in the case of the reception of the synaptic weights and the reception of the second sample). Alternatively, it may be a wired communication when connecting the motor vehicle to an electric charging station, in the case where the motor vehicle is electrically or hybrid-powered.

[0016] According to one embodiment, the first artificial neural network comprises a first recurrent neural network with short-term and long-term memory.

[0017] For example, the first artificial neural network consists of: • A dense layer of neurons, one output of which is connected to the output of the first artificial neural network, and • A recurrent neural network with short-term and long-term memory, completely connected at the output to the dense layer, and one input of which is connected to the input of the first artificial neural network, and

[0018] For example: • The short-term and long-term memory recurrent neural network classically comprises a chain of neural modules, each neural module in the chain: • Receiving, as input to each neuron module, a value from the first series if this value is included in the first sample or the second sample, a neutral value otherwise (for example, equal to zero), and • Being connected, at the output of said each neuron module, to each neuron the dense layer.

[0019] Of course, as a variant, other types or architectures of neural networks are conceivable.

[0020] For example, the neuron module chain comprises 60 modules, and the first series and / or the second series consists of 60 values.

[0021] For example, the driving parameter is a speed of a propulsion engine of the motor vehicle (for example in revolutions per minute).

[0022] Alternatively, it is a speed or an acceleration of the motor vehicle. It can also be other parameters such as the location of the motor vehicle or its energy consumption.

[0023] According to one embodiment, the first artificial neural network comprises neural layers, connections between neurons (and possibly one or more loops), a combination function and an activation function, and the computer processing method further comprises the following fourth steps, implemented by the computing device: • Reception of synaptic weights (following the first transmission or the second transmission), from the motor vehicle after the second stages (and, for example, before or after the third stages), • Memorization (such a memorization step, for a neural network, is known to those skilled in the art) of a second artificial neural network (in other words: of a second electronic neural network) by the computing device, the second artificial neural network comprising (and the method may comprise a step of receiving these elements, alternatively, they are constructed by the computing device): • Neural layers identical to the neural layers of the first artificial neural network (memorization of the neural layers of the second neural network can take place after or before the reception of synaptic weights or the second steps), and • Connections between neurons (and said one or more possible loops) identical to the connections between neurons of the first artificial neural network (the memorization of the connections between neurons of the second neural network can take place after or before the reception of the synaptic weights or the second steps), and • A combination function identical to the combination function of the first artificial neural network (the storage of the combination function of the second neural network can take place after or before the reception of the synaptic weights or the second steps), and • An activation function identical to the activation function of the first artificial neural network (memorization of the activation function of the second neural network can take place after or before receiving the synaptic weights or the second steps), and • Synaptic weights identical to the synaptic weights of the first artificial neural network, memorized after reception of the synaptic weights, • An output identical to the output of the first artificial neural network, and • An input identical to the input of the first artificial neural network. • Receiving the second sample, from the motor vehicle, for example, via a radio frequency communication means, after the second steps (and, for example, before or after the third steps), • Then, production (in other words: prediction), by the second artificial neural network, of a fourth series, at the output of the second artificial neural network, from the second sample, (received) at the input of the second artificial neural network.

[0024] According to one embodiment, the method may then comprise a step of analysis (in other words: processing) of the fourth series, for example, to program or carry out preventive maintenance steps (for example, to choose parts to be changed) of the first motor vehicle (and the method may comprise these analysis and / or maintenance steps).

[0025] The invention also relates to a computer program comprising instructions executable by a microprocessor or a microcontroller or a computer, for implementing the steps of the method according to the invention, when it is executed by the microprocessor or the microcontroller or the computer.

[0026] The method according to the invention can be implemented by an electronic device (comprising, for example, a microprocessor). The invention therefore also relates to an electronic device (and / or a motor vehicle) configured to implement the steps of the method according to the invention. The invention also relates to a system comprising the motor vehicle and a computer device configured to implement the fourth steps.

[0027] The characteristics and advantages of the computer program, the electronic device, and the system are identical to those of the method according to the invention (without it being necessary to repeat them here).

[0028] When the electronic device, the first motor vehicle, the computer device, or another element is “configured to” (or “capable of”) performing or implementing a step or operation, this implies, for example, that the element comprises means for performing the step or operation. The means preferably comprise electronic means, for example a computer program, data in memory, specialized electronic circuits, a data bus, a microprocessor and / or a microcontroller.

[0029] Other characteristics and advantages of the present invention will appear more clearly on reading the detailed description which follows, comprising embodiments of the invention given as non-limiting examples and illustrated by the appended drawings, in which:

[0030] [Fig-1] represents an electronic device and a motor vehicle, according to a embodiment of the invention, in top view,

[0031] [Fig.2] represents a neural network during the implementation of the method according to the embodiment of [Fig.3].

[0032] [Fig.3] represents an implementation of the method according to the invention, according to an exemplary embodiment, by the electronic device of [Fig.l],

[0033] In [Fig.l], certain elements are, of course, seen through transparency.

[0034] Detailed description of an exemplary embodiment of the invention, with reference to Figures 1 to 3.

[0035] [Fig.l] represents a motor vehicle 100. The motor vehicle 100 comprises a microprocessor 110 storing or implementing a neural network 111. In addition, the microprocessor 110 is connected to the following elements of the motor vehicle 100: a module 120 capable of communicating via a mobile telephone network 300, an engine (for propulsion of the motor vehicle 100) 130 comprising a device (or sensor) for measuring the engine speed 131 of the engine 130.

[0036] In step S00, when the motor vehicle is moving, the microprocessor 110 receives, from the engine speed measuring device 131, a first series of engine speeds 130, for example in revolutions per minute, for example every second, in the chronological order of the measurements made by the engine speed measuring device 131. The second series serie2, [Fig.2], can in particular serve as an example for the first series. It is in this chronological order.

[0037] In step S10, the microprocessor 110 determines a first sample of the first series, in the same manner as in step S60 below.

[0038] In step S20, the microprocessor 110 implements, in a known manner, supervised training of the neural network 111, from the first sample labeled by the first series.

[0039] In a known manner, the neural network 111 comprises neuronal layers, connections between neurons, and possibly a connection loop or connection loops, a combination function, an activation function and synaptic weights, and the training at step S20 modifies the synaptic weights.

[0040] Steps S00 to S20 may be repeated, for example in step S30, for example 300 times, with different first series.

[0041] In step S40, the microprocessor 110 transmits, via the module 120 and the mobile telephone network 300, the synaptic weights of the neural network 111 to a server 200, which receives them.

[0042] The server 200 comprises a previously stored neural network 201.

[0043] The neural network 201 thus comprises: • Neural layers identical to the neural layers of the neural network 111, and • Connections between neurons (and any possible loops) identical to the connections between neurons in the neural network 111, and • A combination function identical to the combination function of the neural network 111, and • An activation function identical to the activation function of the neural network 111, and • Synaptic weights received at step S40. • An output identical to the output of the neural network 111, and • An input identical to the input of the neural network 111.

[0044] In step S50, the microprocessor 110 receives a second series serie2, represented [Fig.2], of engine speeds v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12 in the same manner as in step S00.

[0045] In step S60, the microprocessor 110 determines a second sample of the second series serie2, for example consisting of the regimes v1, v5, v9, by systematically taking the regimes which have these ranks (1, 5, 9) in the second series. For example, the regime v5 of the second sample is preceded by the regimes v2, v3, v4 and followed by the regimes v6, v7, v8 in the second series serie2, the regimes v2, v3, v4, v6, v7, v8 being excluded from the second sample.

[0046] In step S70, the neural network 111, implemented by microprocessor 110, produces a third series, represented [Fig.2], at the output si of the neural network 111, from the second sample, in other words, from a representation of the second sample (which amounts to the same thing) at the input el of the neural network 111.

[0047] For example, the neural network 111 has the architecture shown in [Fig.2], in which the neural network 111 consists of: • A dense neural layer 1 connected to the output si of the neural network 111, and • From a fully connected short-term and long-term memory recurrent neural network Istml at output (for ease of reading, the connections have not been shown), to the dense layer densel, and connected, as input, to the input el of the neural network 111.

[0048] For example, the short-term and long-term memory recurrent neural network Istml comprises a chain of neural modules, each neural module ml, m2, m3, m4, m5, m6, m7, m8, m9, ml0, mil, ml2 of the chain: • Includes, as input, a regime from the second series serie2 if this regime is included in the second sample, a neutral value otherwise (for example, equal to zero), and • Is connected, at the output of said each neuron module ml, m2, m3, m4, m5, m6, m7, m8, m9, ml0, mil, ml2, to each neuron the dense layer dense1.

[0049] Alternatively, of course, other architectures for the Neural Network 111, more or less distant from that presented in [Fig.2], are possible. For example, in the example, [Fig.2], to facilitate reading of the figure, the series serial 1 comprises 12 regimes and the network of recurrent neurons with short-term and long-term memory Istml comprises a chain of 12 modules, but in practice, the series serial could comprise 60 regimes and the network of recurrent neurons with short-term and long-term memory Istml could comprise a chain of 60 modules.

[0050] In step S90, the microprocessor 100 checks whether a distance, calculated in step S80, between the second series serie2 and the third series at output si, is greater than a threshold, for example equal to 1000. For example, in the case of a Euclidean distance, in the example of [Fig.2], the distance between the second series serie2 and the third series, at output si, is equal to 30.

[0051] If the distance is greater than the threshold: • In step S100, the microprocessor 110 implements, in a known manner, supervised training of the neural network 111, from the second sample labeled by the second series serie2 (NB: this step can possibly be repeated), then • In step S110, the microprocessor 110 transmits, via the module 120 and the mobile telephone network 300, the synaptic weights of the neural network 111 (modified in step S100) to the computer device 200, which receives them and replaces the weights of the neural network 201 with those received.

[0052] In step S120, the microprocessor 110 transmits the second sample to the server 200 which receives it, via the module 120 and the mobile telephone network 300.

[0053] Alternatively, steps S70 to S110 are omitted from the method [Fig.3].

[0054] In step S130, the neural network 201, implemented by the server 200, produces a fourth series, at the output of the neural network 201, from the second sample, in other words, from a representation of the second sample (which amounts to the same thing) at the input of the neural network 201.

[0055] In step S140, the server analyzes the fourth series, for example, to program or carry out preventive maintenance steps of the motor vehicle.

Claims

Claims

1. A computer processing method comprising a repetition of the following first steps: • Reception (S00) of a first series of values ​​of a driving parameter of a motor vehicle (100), in a first chronological order, • Determining (S 10) a first sample of the first series of values, the first sample comprising a plurality of first values ​​of the first series of values, in the first chronological order, each value of the plurality of first values ​​being preceded by previous first values ​​in the first series of values, and followed by next first values ​​in the first series of values, the previous first values ​​and the next first values ​​being excluded from the first sample, • First supervised training (S20) of a first artificial neural network (111) comprising an input (el) and an output (si), from the first sample, at the input (el), the first sample being labeled by the first series of values, the first artificial neural network (111) comprising synaptic weights, the first supervised training (S20) modifying the synaptic weights.

2. Computer processing method according to the preceding claim, further comprising the following second steps, implemented in the motor vehicle (100), after repetition of the first steps: • First transmission (S40) of synaptic weights to a computer device (200), then • Reception (S50) of a second series of values ​​(series2) of the driving parameter of the motor vehicle (100), in a second chronological order, • Determination (S60) of a second sample of the second series of values ​​(series2), the second sample comprising a plurality of second values ​​of the second series of values ​​(series2), in the second chor-

3.

4. nological, each value (v5) of the plurality of second values ​​being preceded by previous second values ​​(v2, v3, v4) in the second series of values ​​(series2), and followed by next second values ​​(v6, v7, v8) in the second series of values ​​(series2), the previous second values ​​(v2, v3, v4) and the next second values ​​(v6, v7, v8) not belonging to the second sample, • Transmission (S 120) of the second sample to the computing device (200). Computer processing method according to the preceding claim further comprising the following third steps, implemented, in the motor vehicle, after the step of determining the second sample and before the transmission of the second sample: • First production (S70), by the first artificial neural network (111), of a third series, at the output (si) of the first artificial neural network (serial), from the second sample, at the input (el) of the first artificial neural network (111), • Check (S 80) if a distance between the second series and the third series is greater than a threshold, • If the said distance is greater than the threshold: • Second training (S 100), supervised of the first artificial neural network (111) from the second sample, at the input (el) of the first artificial neural network (111), labeled by the second series, the second training modifying the synaptic weights of the first artificial neural network (111) then, • Second transmission (SI 10) of synaptic weights to the computer device (200). A computer processing method according to any one of the preceding claims wherein the first artificial neural network (111) comprises a short-term and long-term memory (Istml) recurrent neural network.

5. Computer processing method according to any one of the preceding claims in which the driving parameter is a speed of a propulsion engine (130) of the motor vehicle (100).

6. A computer processing method according to any one of the preceding claims taken as a dependency of claim 2 or 3, the first artificial neural network (111) comprising neural layers, connections between neurons, a combination function and an activation function, the computer processing method further comprising the following fourth steps, implemented by the computing device: • Receiving the synaptic weights, from the motor vehicle (100), after the second steps • Storing a second artificial neural network (201) by the computing device (200), the second artificial neural network (201) comprising: • Neural layers identical to the neural layers of the first artificial neural network (111), • Connections between neurons identical to the connections between neurons of the first artificial neural network (111),• A combination function identical to the combination function of the first artificial neural network (111), • An activation function identical to the activation function of the first artificial neural network (111), • Synaptic weights identical to the synaptic weights of the first artificial neural network (111), stored after receiving the synaptic weights, • An output identical to the output (si) of the first artificial neural network (111), and • An input identical to the input (el) of the first artificial neural network (111), • Receiving the second sample, from the motor vehicle (100), after the second steps, • Then, production (S 130), by the second artificial neural network (201), of a fourth series, at the output of the second artificial neural network, from the second sample, at the input of the second artificial neural network (201).,

7. Computer program comprising instructions, executable by a microprocessor or a microcontroller, for implementing the method according to any one of claims 1 to 5, when executed by the microprocessor or the microcontroller.

8. Electronic device (110) configured to implement the steps of the method according to any one of claims 1 to 5.

9. Motor vehicle (100) comprising the electronic device (110) according to the preceding claim.

10. A system comprising the motor vehicle (100) of claim 9 and a computing device (200) configured to implement the fourth steps of claim 6.

Images