vlrPhone |

With VLC and VLR codecs (compression and decompression methods), it is possible to make the connected objects more intelligent, more recognizable and more predictable.

The need to compress at the object level is not obvious, especially if the flows are low. But at the level of the machines that record the data of a few objects to several thousand of objects, the problem of compression can arise, especially over long periods.

For IoT (Internet of Things) and M2M (Machine to Machine), at the level of the machines, the VLC and VLR codecs can provide very important features in artificial intelligence and deep learning for the detection of the anomalies, the recognition of the objects and the prediction of the behaviors.

Internet of Things.

Machine to Machine Communication.

k-Nearest Neighbor.

Approximate Nearest Neighbors.

Artificial Neural Network.

MultiLayer Perceptron.

Convolutional Neural Networks.

Recurrent Neural Network.

Natural Language Processing.

Mel-Frequency Cepstral Coefficients.

Message Queuing Telemetry Transport.

A multilayer perceptron is a class of feedforward artificial neural network. An MLP consists of at least three layers of nodes. Except for the input nodes, each node is a neuron that uses a nonlinear activation function. MLP utilizes a supervised learning technique called backpropagation for training. Its multiple layers and non-linear activation distinguish MLP from a linear perceptron. It can distinguish data that is not linearly separable.

In machine learning, a convolutional neural network (CNN, or ConvNet) is a class of deep, feed-forward artificial neural network, inspired by biological processes in which the connectivity pattern between neurons is inspired by the organization of the animal visual cortex.

Unlike MLP, CNN shares weights in convolutional layers, which means that the same filter (weights bank) is used for each receptive field in the layer; this reduces memory footprint and improves performance.

A recurrent neural network is a class of artificial neural network where connections between units form a directed cycle. This allows it to exhibit dynamic temporal behavior. Unlike feedforward neural networks, RNNs can use their internal memory to process arbitrary sequences of inputs. This makes them applicable to tasks such as unsegmented, connected handwriting recognition or speech recognition.

Natural language processing is a field of computer science, artificial intelligence and computational linguistics concerned with the interactions between computers and human (natural) languages, and, in particular, concerned with programming computers to fruitfully process large natural language corpora. Challenges in natural language processing frequently involve natural language understanding, natural language generation (frequently from formal, machine-readable logical forms), connecting language and machine perception, dialog systems, or some combination thereof.

For a quick description of the VLC and VLR codecs, see at the following address:

Additional Notes:

- The sampling rates should be regular. If this is not the case, interpolations must be done so that the samples are regularly spaced in time.

- The treatments can be done like in audio by playing on the size of the FFT buffers and the number of frames per second.

With audio, one can have FFT buffer sizes of 512 or 1024, with 31.25 frames per second. With the connected objects, one can have one frame every two seconds, one frame per second, 10 frames per second, ...

- It should be noted that there are protocols such as MQTT that can link the connected objects to the machines. The input data may originate from these protocols or may come from files.

- The VLC and VLR codecs contain the magnitudes of the points, the phases of the foreground points as well as the positions of the foreground points and the background bands.

The vectors of the magnitudes and the vectors of the positions or the vectors of the magnitudes and positions directly provide the features for the algorithms of the artificial intelligence and the deep learning.

- In order to minimize the number of points taken into account, it is better to take the greatest points and the most energetic bands considering only the local peaks. The information of the highest importance is kept, without getting bogged down with information of second importance.

This is well suited to all signals, except for signals that contain only essentially white noise.

- Methods such as RNN are particularly suitable for the analysis of the time series.

The magnitudes considered by the VLC and VLR codecs constitute the spectral envelope of the signal. We can simulate the time series by modeling the spectral envelope by interpolation curves, notably by cubic splines.

These algorithms can then be considered as "guided", in addition to being supervised or unsupervised.

Methods such as NLP allow the voice recognition. They are based mostly on MFCC.

In the frequency domain, one moves on to the MEL scale using the critical bands, one takes the logarithm of the magnitudes, one performs DCTs (Discrete Cosine Transform) on the signal (very reduced signal compared to the original, at the MEL scale) and finally one takes a few points as a digital fingerprint.

The VLC and VLR codecs are fully compatible with this approach. The transmitted magnitudes can be considered as quasi-fingerprints. It is enough to make DCTs on these magnitudes (after the logarithm if one compresses the magnitudes otherwise).

In addition, these codecs are not dependent on the psychoacoustic or critical bands. One can take bands adapted to its problem.

The input data can be homogeneous (eg voice, heart rate or temperature).

Our codecs take into account the multichannel (including many channels) with homogeneous data.

Our codecs can also take into account non-homogeneous multichannel data (eg wind pressure, temperature and humidity). One must use the same sampling rate and the same FFT buffer size.

The composite features are to be used with artificial intelligence and deep learning algorithms.

Apart from audio, our codecs can be used in many areas such as the finance, the health or the weather forecasting.

We have already presented two possible projects in the field of health:

vlrMemo

Telemonitoring

We will present here two examples of composite features, the first in the field of the health and the second in the field of the weather forecasting, including the local forecasts, in the short, medium and long term.

All the features share the same scale of positions (frequencies). Some signals may be square (constant values per interval), in the frequency domain, they will be represented by harmonics, therefore only local peaks.

In the field of the health, the data below can provide composite features for estimating and predicting the overall health status:

- The RR Intervals.

- The Heart Rate.

- The Systolic Blood Pressure.

- The Diastolic Blood Pressure.

- The Mean Blood Pressure.

- The Pulse Pressure.

- The Stroke Volume.

- The Cardiac Output.

- The Left Ventricular Ejection Time.

- The Total Peripheral Resistance.

- The Aortic Characteristic Impedance.

- The Arterial Compliance.

The meteorological data can be compressed particularly efficiently with the VLC and VLR codecs, because of their repetitive patterns.

Short, medium and long term local forecasts, based on artificial intelligence and deep learning are possible, using the composite features, for example:

- The temperature.

- The minimum temperature.

- The maximum temperature.

- The atmospheric pressure.

- The speed of the wind.

- The direction of the wind.

- The humidity level.

- The amount of cloud.

- The value of visibility.

- The value of precipitation.

- The volume of rain.

- The volume of snowfall.

- The global solar radiation.

- The direct solar radiation.

- The diffuse solar radiation.

It seems to us that it is quite possible to use the methods of the VLC and VLR codecs in the case of the spherical harmonics.

In the one-dimensional case, the Fourier coefficients depend on a variable m (the frequency).

In the case of spherical harmonics, the Fourier coefficients depend on two variables n (the order) and m. For each value of n, we have a one-dimensional problem and we can perform decimations (coefficients set to zero).

One can easily increase the order while minimizing the number of features for the artificial intelligence or the deep learning, or while omitting the calculations of parameters related to zero coefficients.

A weight system may be used depending on the frequency (and depending on the time if there are temporal movements) in order to better choose the points to keep.