SYSTEM FOR PROCESSING DEFORMATION MEASUREMENTS OF A SILO SUPPORT FOOT FOR ISOLATING THE DEFORMATION COMPONENT RELATED TO THE WEIGHT OF THE SILO CONTAINER THROUGH AUTOMATIC LEARNING

Abstract

The invention relates to a system for processing deformation measurements of a storage silo support base, which improves upon the system described in EP3963296B1. The system includes deformation sensors and sensors for detecting transient environmental phenomena associated with the support base. A remote server integrates a specific machine learning model, pre-trained for each sensor, which generates time series of deformation measurements sequentially or in real time. These deformation components make it possible to isolate the effects of temperature and sunlight variations to improve the accuracy of determining the silo's fill level. This solution optimizes the structural monitoring of storage silos by ensuring greater measurement reliability, without requiring manual recalibration when environmental conditions change. Figure to be published with the abstract: 2

Description

Title of the invention: SYSTEM FOR PROCESSING DEFORMATION MEASUREMENTS OF A SUPPORT FOOT SILO FOR ISOLATING THE DEFORMATION COMPONENT RELATED TO THE WEIGHT OF THE CONTAINER SILO BY AUTOMATIC LEARNING technical field

[0001] The invention relates to the field of structural monitoring systems applied to silos for storing or mixing bulk materials.

[0002] More specifically, the invention relates to systems for processing measurements of deformation of silo support feet, intended to determine the filling level of these structures by analyzing the mechanical stresses they are subjected to. Prior art

[0003] In the silo operation sector, a common problem is to estimate the filling level of a silo.

[0004] This problem is known from document EP3963296B1 which describes a system for monitoring structural deformations of storage silos using a single deformation sensor fixed on a single silo support foot to measure mechanical stresses.

[0005] In this document, the estimation of the filling level of a silo is carried out using a mathematical law which links the deformation values ​​to the filling level.

[0006] Although the influence of transient environmental phenomena on deformation measurements, such as temperature variations or exposure to solar radiation, is a fact known to those skilled in the art, conventional methods for addressing these influences prove insufficient given the complexity of the system. Indeed, the non-uniform temperature distribution, the complex geometry of the structure, and the heterogeneous distribution of mechanical stresses create interactions that cannot be satisfactorily compensated by traditional correction approaches.

[0007] In practice, these environmental phenomena induce deformation components which are superimposed on that related to the weight of the contents, making it more complex to accurately estimate the filling level of the silos.

[0008] Thermal variations, in particular, cause expansion of materials which adds to the deformation component related to the weight of the contents.

[0009] Although the system described in EP3963296B1 allows a reliable estimation of the filling level, it does not offer a solution for effectively isolating the different deformation components.

[0010] This limitation affects the accuracy of current monitoring systems for tracking the filling level of silos.

[0011] Thus, there is a need for a system capable of automatically isolating, in deformation measurements, the component related to the weight of the contents by identifying and eliminating the components due to transient environmental phenomena, in order to improve the accuracy of determining the filling level of storage silos. Summary of the invention

[0012] The invention aims to solve, at least partially, this need.

[0013] A first aspect of the invention relates to a system for processing deformation measurements of a silo support foot of a storage or mixing silo for at least one bulk material, the system being specifically designed to collect the deformation measurements and isolate the component related to the weight of the contents for use in determining the fill level of the silo, the system comprising, - at least one strain sensor that is mechanically coupled to a single silo support foot and that is specifically designed to measure the strain of the silo support foot in response to the introduction or removal of bulk material, the strain sensor comprising at least a first electronic circuit that is specifically designed to generate, at least a first measurement signal that includes strain values ​​and timestamp data that can be integrated into the first measurement signal or transmitted separately, - at least one sensor for detecting transient environmental phenomena that is mechanically coupled to the silo support base or that is located near the silo support base, - at least one initial wireless communication means specifically designed to transmit the initial measurement signal with its associated timestamp data, - at least one remote server comprising, — at least one second wireless communication means specifically designed to receive the first measurement signal with its associated timestamp data, — at least one first processor, and — at least one primary storage memory, in which, - The transient environmental phenomenon detection sensor is specifically designed to measure at least one of the following quantities: an ambient temperature quantity and / or a quantity representative of sunlight conditions. The transient environmental phenomenon detection sensor includes at least one second electronic circuit specifically designed to generate — at least one second measurement signal that includes local environmental measurement data corresponding to the measured quantities and characteristics of transient environmental phenomena affecting the silo support base due to its exposure to local environmental conditions, and — timestamp data that can be integrated into the second measurement signal or transmitted separately, - the first wireless communication method is also specifically designed to transmit the second measurement signal with its associated timestamp data, - The second wireless communication method is specifically designed to receive the second measurement signal with its associated timestamp data; - The first processor is specifically designed to, — for strain sensors, apply at least one machine learning model, previously trained to process data from one or more strain sensors, to generate time series describing the strain variations linked to environmental phenomena according to the installation conditions on the silo base, — the machine learning model receiving as input at least one matrix comprising measurement signal data with their associated timestamp data, data including deformation values ​​and their associated timestamps, and environmental values ​​and their associated timestamps, — for each strain sensor, sequentially or in real time exploit the time series generated by its specific machine learning model to isolate the different strain components, based on the corresponding timestamps, - the first storage memory is specifically designed to store, for each strain sensor, — its specific pre-trained machine learning model, and — the time series describing the deformation variations generated by its specific machine learning model.

[0014] In a first embodiment of the first aspect of the invention, the first processor is specifically designed to apply time series describing the deformation variations generated by the machine learning model according to the following steps, - categorize the local environmental measurement data extracted from the second measurement signal into predetermined time periods corresponding to predetermined ranges of variation, - determine categorization thresholds based on, — the statistical distribution of local environmental measurement data in the training dataset, and / or — transient environmental phenomena affecting the silo support base, - for each predetermined time period, — determine at least one statistically representative value of the deformation components for the period based on local environmental measurement data extracted from the second measurement signals with their associated timestamp data, — store, in the first storage memory, the statistical value in association with the local environmental measurement data and their timestamps, — generate deformation components specific to the position of the deformation sensor on the silo support foot using the statistical value representative of the deformation components for the corresponding time period.

[0015] In a particular implementation of the first embodiment of the first aspect of the invention, the predetermined ranges of variation comprise: - a range of small variations, below a first predetermined threshold, - a range of median variations, between the first threshold and a second predetermined threshold, and - a range of high variations, above the second predetermined threshold, in which the first and second thresholds are determined as categorization thresholds.

[0016] In a second embodiment of the first aspect of the invention, the first processor is specifically designed to apply sequentially or in real time the process of isolating the deformation components to the deformation measurements according to the following steps, - for each new deformation measurement received at a given time via the first measurement signal with its associated timestamp data, — determine a range of variation for local environmental measurement data over a predetermined time period including the last measurement instant, — select, from the time series generated by the specific machine learning model, a time series corresponding to the measurement data local environmental data extracted from the second measurement signal with its associated timestamp data for the range of variation, - extract from the time series a deformation component corresponding to, — the time of the new deformation measurement determined from the timestamp data associated with the first measurement signal, and — local environmental measurement data extracted from the second measurement signal with its associated timestamp data for the determined range of variation, - apply the strain component to the strain measurement to obtain an isolated measurement.

[0017] In a third embodiment of the first aspect of the invention, during its use, the machine learning model is specifically designed to: - receive as input at least one matrix comprising, — previous strain values ​​with their associated timestamp data, and — at least one exogenous variable corresponding to local environmental measurement data with their associated timestamp data, the exogenous variable being used as an indicator of transient environmental phenomena, - generate a time series describing the expected deformation variations under the environmental conditions corresponding to the input matrix.

[0018] In a fourth embodiment of the first aspect of the invention, when using the machine learning model, - The sensor for detecting transient environmental phenomena is specifically designed to generate at least two distinct exogenous variables corresponding respectively to, — initial local environmental measurement data extracted from the second measurement signal with its associated timestamp data, and — second set of local environmental measurement data extracted from the second measurement signal with its associated timestamp data, characteristic of seasonal cycles, - the machine learning model is specifically designed for, — receive as input at least one matrix comprising the first and second data points, and — preferably generate a single time series describing the deformation variations linked to environmental phenomena corresponding to the deformation values, environmental data and their respective timestamps contained in the input matrix including the first and second data, or alternatively, generate several distinct time series based on these same data.

[0019] A second aspect of the invention relates to a training system for at least one machine learning model for processing deformation measurements of a silo support foot of a storage or mixing silo for at least one bulk material, the machine learning model(s) being configured to process deformation measurements from one or more sensors installed on the silo support foot, the system being specifically designed to generate deformation components adapted to the installation conditions of the sensors for use in determining the fill level of the silo, the system comprising, - at least one second processor specifically designed to, — identify time periods during which the deformation variations are mainly due to transient environmental phenomena, including,— the analysis of deformation variations from measurements acquired by at least one deformation sensor that is mechanically coupled to the silo support base, the deformation sensor being configured to generate at least one first measurement signal comprising deformation values ​​and timestamp data that can be integrated into the first measurement signal or transmitted separately, — the detection of regular variation patterns corresponding to measurements acquired by at least one transient environmental phenomenon detection sensor that is mechanically coupled to the silo support base or that is located near the silo support base, the sensor being configured to generate at least one second measurement signal comprising local environmental measurement data and timestamp data that can be integrated into the second measurement signal or transmitted separately, — the creation of a training dataset comprising,— Extracting deformation data and their associated timestamps from measurements acquired by the deformation sensor, — Centering the deformation data to eliminate time lags, — Concatenating the data while preserving temporal consistency, — Training the machine learning model on the dataset to generate time series describing deformation variations, including, — Preparing at least one input matrix comprising data extracted from the measurement signals with their associated timestamps, data including deformation values ​​and their associated timestamps, and environmental values ​​and their associated timestamps, — Using the input matrix to train the model to generate deformation component predictions taking into account, — transient environmental phenomena affecting the silo support base due to its exposure to local environmental conditions according to its specific position, and — specific mechanical stresses exerted on the sensor due to its position on the silo support structure, - at least one second storage memory specifically designed to store — the training dataset, — the machine learning model, and — the time series generated by the machine learning model.

[0020] In a first embodiment of the second aspect of the invention, the second processor is specifically designed to identify time periods according to at least one of the following approaches: - detection of periods of silo inactivity, - detection of regular and cyclical variation patterns characteristic of transient environmental phenomena, and - analysis of deformation variations correlated with local environmental measurement data, in which the time periods include periods of at least consecutive hours during which the silo exhibits regular variations in strain sensor measurements corresponding to seasonal cycles.

[0021] In a second embodiment of the second aspect of the invention, the second processor is specifically designed to: - categorize the local environmental measurement data extracted from the second measurement signal into predetermined time periods corresponding to predetermined ranges of variation; - determine categorization thresholds based on: - the statistical distribution of the local environmental measurement data in the training dataset; and — transient environmental phenomena affecting the silo support base, - for each predetermined time period, — determine at least one statistically representative value of the deformation components for the period, — store, in the second storage memory, the statistical value representative of the deformation components for the corresponding time period in association with the local environmental measurement data and their timestamps.

[0022] In a particular implementation of the second embodiment of the second aspect of the invention, the predetermined ranges of variation include, - a range of small variations, below a first predetermined threshold, - a range of median variations, between the first threshold and a second predetermined threshold and - a range of high variations, above the second predetermined threshold, in which the first and second thresholds are determined as categorization thresholds.

[0023] In a third embodiment of the second aspect of the invention, the second processor is specifically designed to periodically update the machine learning model according to the following steps: - detection of a significant change in local environmental measurement data over a predetermined period corresponding to a seasonal variation, including: - calculation of a difference between the current local environmental measurement data and that used for training the machine learning model, taking into account the associated timestamps; - comparison of the difference to a predetermined threshold; - when the difference exceeds the predetermined threshold, - automatic creation of a new training dataset comprising the deformation values ​​and their timestamps, and the local environmental measurement data and their timestamps.— retraining the machine learning model with the new dataset, and , — generation of new time series.

[0024] In a fourth embodiment of the second aspect of the invention, the machine learning model is a time series prediction model specifically designed to: - receive as input at least one matrix comprising: - deformation values ​​with their associated timestamp data, and - at least one exogenous variable corresponding to local environmental measurement data with their associated timestamp data, the exogenous variable being used as an indicator of transient environmental phenomena; - generate a time series describing the expected deformation variations under the environmental conditions corresponding to the input matrix.

[0025] In a particular implementation of the fourth embodiment of the second aspect of the invention, the generation of the time series according to a two-phase process comprising, — a first phase of creating and training a distinct model for each category of values ​​of the exogenous variable, and — a second phase of sequential or real-time exploitation of pre-trained models, where the system identifies the category of the current exogenous variable and uses the corresponding model to generate a time series describing the shape of deformation variations linked to environmental phenomena under these specific conditions.

[0026] In a fifth embodiment of the second aspect of the invention, the transient environmental phenomenon detection sensor is specifically designed to generate at least two distinct exogenous variables corresponding respectively to: first, local environmental measurement data extracted from the second measurement signal with its associated timestamp data, and — second set of local environmental measurement data extracted from the second measurement signal with its associated timestamp data, characteristic of seasonal cycles, - the machine learning model is specifically designed to — receive as input at least one matrix comprising the first and second data points, and — generate separate time series based on the input matrix comprising the first and second data points. Brief description of the drawings

[0027] Other features and advantages of the invention will be better understood from the following description and with reference to the accompanying drawings, given by way of illustration and not limitation.

[0028] [Fig-1] Fig. 1 represents a deformation measurement processing system of a storage silo support foot with isolation of deformation components by machine learning, according to the invention.

[0029] [Fig.2] The [Fig.2] represents a schematic block view of the system of the [Fig.1].

[0030] [Fig.3] Fig.3 represents a schematic block view of a system training of a machine learning model for the isolation of the deformation components of a silo support foot, according to the invention.

[0031] The figures do not necessarily respect the scales, for illustrative purposes only. Description of the implementation methods

[0032] Preliminary remarks

[0033] In order not to obscure the description and distract the reader from understanding the principles of the invention, our explanations will not go beyond what we consider necessary for a person in the technical field of material handling equipment to understand and appreciate the underlying concepts of the invention. Indeed, the embodiments illustrated in the description are, for the most part, composed of elements known to a person skilled in the art.

[0034] Technical terms specific to the invention are defined as they are introduced in the description. These definitions are essential for a precise understanding of the invention. The reader is encouraged to refer to them regularly to correctly interpret the scope and operation of the invention. Generally, each definition is presented immediately after the first occurrence of the corresponding term.

[0035] Objective of the invention

[0036] One of the main objectives of the invention is to meet the identified need by providing a system capable of automatically isolating, in the deformation measurements of a silo support foot, the component related to the weight of the contents. This isolation is achieved by identifying and quantitatively separating the components due to transient environmental phenomena, thus allowing for a more precise analysis of the deformation related to the weight of the contents.

[0037] The term "isolate" in the context of the invention refers to the mathematical and technical process of decomposing a total deformation signal into its various constituent components. This operation makes it possible to identify and quantitatively separate each specific contribution to the overall measured deformation. For example, when the system isolates the component related to the weight of the contents, it performs a mathematical analysis that decomposes the total deformation signal to extract only the portion corresponding to the mechanical stresses induced by the mass of the stored material, eliminating the effects of other factors such as temperature or sunlight. This isolation relies on machine learning techniques that make it possible to recognize and quantify the different deformation components according to their specific characteristics and their correlation with the measured environmental variables.

[0038] To this end, the inventors propose an improvement of the system described in EP3963296B1 by integrating sensors for detecting transient environmental phenomena mechanically coupled to the silo support foot or placed near the silo support foot, associated with machine learning processing to identify and isolate the different deformation components.

[0039] In particular, the system uses one or more machine learning models to process data from strain sensors. Depending on the installation characteristics and monitoring requirements, the system can employ either a single model to process data from several sensors with similar conditions, or separate models for sensors requiring individualized processing. This flexible approach optimizes processing by taking into account positional and environmental exposure characteristics, whether shared by a group of sensors or specific to a given sensor.

[0040] This approach makes it possible to overcome the limitations of the EP3963296B1 system by isolating in the deformation measurements the component related to the weight contained in the silo, thus significantly improving the accuracy of the determination of the filling level of the silo.

[0041] First aspect of the invention: a system for processing deformation measurements of a silo support foot of a storage silo with isolation of the deformation components by machine learning

[0042] As illustrated in [Fig.1] and [Fig.2], a first aspect of the invention relates to a system 100 for processing deformation measurements of a silo support foot 11 of a silo 10 for storing or mixing at least one bulk material, which is specifically designed for isolating the deformation components related to the weight of the contents.

[0043] The term "strain measurement" refers to the process of quantitatively assessing the deformation undergone by a structural element, such as a silo support base 11, under the effect of mechanical stresses. This measurement is generally carried out using specific sensors, such as strain gauges, which convert the mechanical deformation into a usable electrical signal. For example, strain measurement can be used to continuously monitor the evolution of the elongation or compression of a silo base as a function of the fill level. It can also be used to detect any structural anomalies or degradation by comparing the measured values ​​to predefined thresholds. Furthermore, strain measurement can be used as input data for predictive models designed to estimate the stresses experienced by the structure and to anticipate the necessary maintenance operations.

[0044] As illustrated in [Fig. 1], the term "support foot" refers to a vertical structural element on which a storage silo rests and which transfers the forces to the foundations. Support feet are generally made of steel and dimensioned to withstand the static and dynamic forces induced by the weight of the silo and its contents. By way of illustration, support feet may take the form They consist of square or circular metal posts, fitted with ground fixing plates. They can also be made of concrete pillars, poured on site or prefabricated, with internal metal reinforcements to increase their strength.

[0045] As illustrated in [Fig. 1], the term "storage or mixing silo for at least one bulk material" refers to an enclosed structure designed to contain and homogenize divided solid materials, such as aggregates, powders, or grains. These silos are generally cylindrical in shape and can be made of steel, concrete, or composite materials. They are equipped with filling, emptying, and mixing systems to facilitate the handling or mixing of the stored products. For example, a storage or mixing silo can be used in a concrete plant to separately store the different components (sand, gravel, cement) and dose them according to the proportions required for each type of concrete. It can also be used in the food processing industry to store and homogenize batches of grain before processing.

[0046] In particular, the system 100 is specifically designed to collect deformation measurements and isolate the component related to the weight of the contents. This isolated component is then used to determine the fill level of the silo.

[0047] The term "isolating the component related to the weight of the contents" refers to the process of identifying and separating the different deformation components measured by the sensors. This process aims to specifically extract the deformation component induced by the weight of the contents stored in the silo, eliminating the effects of variations in temperature, sunlight, and other external factors. This operation makes it possible to obtain measurements that are more representative of the deformation actually caused by loading the silo. Isolating this specific component is essential for an accurate estimation of the fill level, freeing it from environmental disturbances that can affect the raw deformation measurements.

[0048] The expression "for the purpose of determining the silo fill level" indicates that the main objective of processing the deformation measurements is to isolate the component related to the weight of the contents in order to allow for an accurate estimation of the quantity of material stored in the silo. Indeed, this isolated component has a mathematical relationship with the total weight of the contents, which can be converted into a volume or mass of material, taking into account its density. This relationship can be linear, polynomial, or follow other mathematical models depending on the geometry of the silo, the characteristics of the stored material, and the sensor installation conditions. As an example, this isolated component can be used as an input variable in a mathematical model that establishes a relationship between the load applied to the feet and the corresponding fill level, based on the The geometry of the silo and the properties of the stored material are taken into account. It can also be used to calibrate and verify readings from other types of level sensors, such as radar probes or integrated weighing systems. Furthermore, continuous monitoring of the fill level from this isolated component can optimize inventory management by automatically triggering alerts or replenishment orders when certain thresholds are reached.

[0049] As illustrated in [Fig.1] and [Fig.2], the system 100 comprises at least one strain sensor 110, at least one transient environmental phenomenon detection sensor 120, at least one first wireless communication means 130 and at least one remote server 140.

[0050] - the deformation sensor

[0051] In the invention, the strain sensor 110 is mechanically coupled to a single silo support foot 11.

[0052] The term "mechanically coupled" means that two elements are physically connected in such a way that forces and movements can be transmitted directly from one to the other without interruption. This mechanical coupling allows the strain gauge to accurately measure the total deformation of the structure to which it is attached, including all deformation components. By way of illustration, a strain gauge can be glued or welded to the surface of a metal support foot, so that any elongation or compression of the foot results in a proportional deformation of the gauge. Mechanical coupling can also be achieved by specific fastening devices, such as flanges or brackets, which keep the gauge in close contact with the structural element while allowing it to follow its movements.Furthermore, in some cases, the sensor can be directly integrated into the material constituting the structure, for example by embedding an optical fiber in the concrete of a support foot to measure variations in internal deformation.

[0053] In addition, the strain sensor 110 is specifically designed to measure the deformation of the silo support foot 11 in response to the introduction or extraction of bulk material.

[0054] Furthermore, the strain sensor 110 includes at least a first electronic circuit (not shown) specifically designed to generate at least a first measurement signal 111 which includes strain values.

[0055] The term "measurement signal" refers to a physical quantity carrying information, emitted by a sensor to transmit the values ​​of the measured quantity. It is generally an electrical signal, whose characteristics (amplitude, frequency, phase, etc.) vary according to the input quantity according to a specific law. For example, a resistive gauge strain gauge produces a signal of The measurement is in the form of an electrical voltage, the value of which is proportional to the relative elongation of the gauge. Similarly, a thermocouple temperature sensor delivers a voltage-type measurement signal, the amplitude of which depends on the temperature difference between the hot and cold junctions. Measurement signals can be transmitted remotely via wired or wireless communication channels for processing and analysis by data acquisition and control systems. Furthermore, they can be combined or multiplexed to allow the simultaneous transmission of several measured quantities on the same medium.

[0056] The term "strain value" refers to a scalar quantity that quantifies the magnitude of the total deformation undergone by a structural element, as measured by a dedicated sensor. This total deformation is the sum of several components: deformation due to the weight of the contents, deformation due to temperature, deformation due to sunlight, and deformation due to other undetermined phenomena.

[0057] In addition, the first electronic circuit generates timestamp data that can be integrated into the first measurement signal 111 or transmitted separately. This timestamp data can also be accompanied by additional information such as diagnostic data or configuration parameters, like their serial number, measurement range, or operating status.

[0058] The term "time stamp" refers to temporal information associated with each measured value or recorded event, which allows them to be precisely located in time. Time stamping is generally achieved by assigning each data point a label indicating the date and time of its acquisition, with a resolution adapted to the dynamics of the observed phenomenon. By way of illustration, for a strain sensor sampled at a frequency of 100 Hz, each measured value can be time-stamped with millisecond precision, making it possible to faithfully track the temporal evolution of stresses. Similarly, for a monitoring system recording discrete events, such as threshold exceedances or changes of state, each occurrence can be time-stamped to the nearest second to facilitate the chronological analysis of incidents.Time stamping ensures data consistency and traceability by allowing correlation and contextualization. It can be performed automatically by acquisition devices using a synchronized reference clock.

[0059] - the sensor for detecting transient environmental phenomena

[0060] In the invention, the transient environmental phenomenon detection sensor 120 can either be located near the silo support foot 11, or mechanically coupled to it.

[0061] The term "transient environmental phenomena" refers to temporary variations in ambient conditions around a structure that induce additional deformation components superimposed on the deformation component related to the weight of the contents. These phenomena can be of natural origin, such as fluctuations in temperature, humidity, or atmospheric pressure related to diurnal or seasonal cycles, or of anthropogenic origin. For example, direct exposure of a metal support leg to solar radiation can cause its thermal expansion and add a thermal deformation component to the measurements, without the actual load borne by the leg changing. Similarly, gusts of wind can cause the structure to vibrate and induce transient deformation components in the sensor signals, without altering the component related to the weight of the contents.Taking into account these transient environmental phenomena makes it possible to isolate the deformation component related to the weight of the contents, by identifying and separating the different deformation components.

[0062] In a first configuration, the transient environmental phenomenon detection sensor 120 is located near the silo support foot 11.

[0063] The expression "in proximity" means that two elements are located close enough to each other to be able to interact or be subject to the same influences, without necessarily being in direct contact. This notion of proximity is relative and depends on the nature and scope of the phenomena considered. By way of illustration, a temperature sensor placed a few centimeters from the support leg of a silo 11 can be considered close enough to measure the thermal variations experienced by the leg, even if it is not in contact with it. Similarly, a weather station located within a radius of a few tens of meters around a silo can provide data representative of the climatic conditions to which the structure is exposed, without having to be installed in its immediate vicinity.

[0064] In a second configuration, the transient environmental phenomenon detection sensor 120 is mechanically coupled to the silo support foot 11, thus allowing a direct measurement of the effects of environmental phenomena on the structure.

[0065] In addition, the transient environmental phenomenon detection sensor 120 is specifically designed to measure at least one of the following quantities: an ambient temperature quantity and / or a quantity representative of sunlight conditions.

[0066] The term "ambient temperature quantity" refers to a quantitative measure of the thermal energy present in the immediate environment of the monitored structure. This quantity is generally expressed in degrees Celsius or Kelvin and reflects The thermodynamic state of the air or environment surrounding the sensors and structural elements. For example, the ambient temperature can be measured by a platinum resistance thermometer attached to the support foot of silo 10, providing a precise indication of the thermal variations to which the structure is subjected. It can also be obtained from a thermocouple integrated into the housing of a strain gauge, thus allowing the thermal component in the measurements to be identified. Furthermore, the ambient temperature can be derived from a network of wireless sensors distributed around the silo, providing a detailed thermal map of the environment.

[0067] The expression "a quantity representative of solar irradiance conditions" refers to a quantitative measurement that characterizes the intensity or duration of exposure to solar radiation of a structure or sensor. This quantity can be expressed in different forms, depending on the type of sensor used and the specific aspect of solar irradiance that one wishes to quantify. By way of illustration, the quantity representative of solar irradiance conditions could be the global irradiance, measured in watts per square meter (W / m²) by a pyranometer installed near the silo. It could also correspond to the daily duration of sunshine, expressed in hours, recorded by a heliograph. In some cases, this quantity can be derived indirectly, for example by measuring the temperature difference between two identical surfaces, one exposed to the sun and the other in the shade, using thermocouples.

[0068] Furthermore, the transient environmental phenomenon detection sensor 120 includes at least one second electronic circuit (not shown) specifically designed to generate at least one second measurement signal 121 which includes local environmental measurement data, corresponding to the measured quantities and characteristics of the transient environmental phenomena contributing to the deformation components of the silo support foot 11 due to its exposure to local environmental conditions, and their associated timestamp data.

[0069] The term “local environmental measurement data” refers to all the quantitative information collected by the transient environmental phenomenon detection sensors 120, which characterize the specific ambient conditions in the immediate vicinity of the monitored structure. This data is considered local because it reflects the particular environment of the site, which may differ from the general meteorological conditions of the region. For example, local environmental measurement data may include the air temperature measured at different heights along the silo support base 11 of the silo 10, the relative humidity near the deformation sensors, or the wind speed and direction at the top of the structure. It may also include measurements of atmospheric pressure, precipitation, or incident solar radiation. In addition, this data may incorporate information on the presence of air pollutants or suspended particles, which could affect the behavior of the sensors or the durability of the materials.

[0070] In addition, the second electronic circuit generates timestamp data and supplementary data that can be integrated into the second measurement signal 121 or transmitted separately. This timestamp data is associated with local environmental measurement data.

[0071] - the first means of wireless communication

[0072] In the invention, the first wireless communication means 130 is specifically designed to emit the first measurement signal 111 with its timestamp data and associated supplementary data.

[0073] The term “wireless communication means” refers to a device or set of electronic components capable of transmitting and / or receiving information without using a direct physical connection, such as cables or optical fibers. These communication means generally use electromagnetic waves to carry data over varying distances, depending on the technology employed. For example, a wireless communication means could be a radio module using the Zigbee protocol, which is particularly well-suited to low-power sensor networks. It could also be a Wi-Fi transceiver integrated into a sensor's housing, enabling data to be transmitted to a central access point. In some cases, the wireless communication means could be a 4G or 5G cellular modem, providing long-range connectivity to transmit data directly to a remote server.In addition, these methods may include technologies such as Bluetooth Low Energy for local sensor configuration, or satellite communication systems for remote sites.

[0074] In addition, the first wireless communication means 130 is also specifically designed to transmit the second measurement signal 121 with its associated timestamp data and supplementary data.

[0075] - the remote server

[0076] In the invention, the remote server 140 comprises at least one second wireless communication means 141, at least one first processor 142, and at least one first storage memory 143.

[0077] The term “remote server” refers to a computer or set of centralized computing resources, physically separate from the sensors and monitored structures, which is responsible for receiving, processing, storing, and analyzing the data collected in the field. This server acts as a central point for all the information and provides the computing power necessary for the processing. complex. For example, the remote server can be a virtual machine hosted in a cloud data center, offering significant flexibility in terms of computing power and storage. It can also be a dedicated physical server, installed on the company's premises, operating the data silos to ensure complete control over sensitive data. In some cases, the remote server may consist of a cluster of geographically distributed machines, ensuring high availability and load balancing. Furthermore, the remote server can integrate automatic backup systems, advanced security mechanisms, and user interfaces for operators.

[0078] - in the remote server: the second wireless communication method

[0079] In the invention, the second wireless communication means 141 is specifically designed to receive the first measurement signal 111 with its associated timestamp data and supplementary data.

[0080] In addition, the second wireless communication means 141 is also specifically designed to receive the second measurement signal 121 with its associated timestamp data and supplementary data.

[0081] - in the remote server: the first processor

[0082] In the invention, the first processor 142 is specifically designed to apply one or more machine learning models to process data from strain sensors 110. The system can use either a single model to process data from several sensors with similar characteristics or installation conditions, or individualized models specifically dedicated to each sensor. This flexible approach is a fundamental aspect of the invention, making it possible to capture the characteristics specific to each measurement point or group of measurement points with similarities.

[0083] The term "processor" refers to an electronic processing unit capable of executing instructions and performing arithmetic and logical operations on received data. In the invention, the processor 142 is specifically designed to apply complex algorithms to deformation measurements and environmental data in order to isolate the different deformation components. By way of illustration, the processor 142 can be a microcontroller embedded in a sensor node, responsible for preprocessing the data before its transmission. It can also be a high-performance multi-core processor 142 integrated into the remote server, capable of running several instances of the machine learning model in parallel. In some cases, the processor 142 can be a graphics processing unit (GPU) specialized in processing the artificial neural networks used for isolating the deformation components.Furthermore, the term 142 processor can refer to a system on a chip (SoC) integrating both . computing units, memory and communication interfaces, optimized for signal processing applications.

[0084] The expression "specific machine learning model" refers to an algorithm or set of mathematical algorithms designed to analyze and interpret deformation and environmental data associated with a particular sensor of the system 100. As a particularly advantageous example, the specific machine learning model can be a Long Short-Term Memory (LSTM) neural network. This type of recurrent neural network is particularly well-suited to processing deformation time series due to its ability to capture long-term dependencies in the data. The LSTM architecture typically comprises two successive LSTM layers, the first with 128 units and the second with 256 units, followed by a dense layer of 512 neurons.The LSTM model proves particularly effective for processing structural monitoring data, especially during extreme weather events such as typhoons, thanks to its ability to simultaneously analyze temporal and spatial correlations between different measurements. This specific machine learning model can learn from historical data and generate time series describing expected deformation variations based on new information received. For example, the specific machine learning model can be a recurrent neural network (RNN) trained to identify deformation components related to variations in temperature and solar radiation. It can also be a set of models, such as a random forest, where each decision tree specializes in identifying a specific deformation component.In some cases, the model may be based on support vector regression (SVR) techniques to capture the nonlinear relationships between the input variables and the deformation components. Furthermore, the specific machine learning model may incorporate attention mechanisms to focus on the most relevant aspects of the input data depending on the context.

[0085] The term "pre-trained" means that the machine learning model has already undergone a process of optimizing its internal parameters from a historical dataset before being deployed to process new measurements. In the case of an LSTM network, this pre-training typically involves optimizing the weights of the entry, forgetting, and exit gates on a dataset covering at least one complete cycle of the periodic variations to be analyzed, in particular the 24-hour daily cycle, with a sampling frequency adapted to the dynamics of the environmental phenomena (e.g., one measurement every 5 minutes). The training generally uses the algorithm of Backpropagation over time (BPTT) with the Adam optimizer and a mean squared error loss function are used. This pre-training allows the model to acquire in-depth knowledge of the relationships between the input variables and the various deformation components, specific to each sensor and its environment. For example, a pre-trained model might have been optimized over several months of data collected during different seasons, enabling it to capture long-term cyclical variations. It might also have been exposed to simulated scenarios of extreme conditions, such as heat waves or storms, to improve its robustness. In some cases, pre-training may include a cross-validation phase to assess the model's generalizability to unseen data.Furthermore, the term may imply that the model has been refined using transfer learning techniques, starting from a generic model pre-trained on a large dataset, and then adapting it to the specific characteristics of the monitored site.

[0086] In practice, the strain sensors 110 are associated with one or more machine learning models, trained to generate time series describing strain variations related to environmental phenomena. A single model can be used to process data from several sensors with similar characteristics or exposure conditions, while individualized models can be developed for sensors requiring specific processing. This flexible approach makes it possible to optimally account for the particular conditions of exposure to environmental phenomena, whether they are common to a group of sensors or specific to a particular sensor.

[0087] The term "time series" refers to ordered sets of numerical values, indexed by time, that represent the expected deformation variations due to transient environmental phenomena under specific conditions. These series are generated by the machine learning model which, having learned the characteristic deformation patterns for different environmental conditions, can identify the deformation component related to environmental phenomena when similar conditions recur. By way of illustration, the model may have learned that, for a given temperature and a specific level of sunlight, the deformation of the metal support base follows a certain temporal profile. Thus, when these same environmental conditions are detected, the system can isolate this component from the total measured deformation.Time series can also capture seasonal cyclical variations, allowing us to anticipate deformations linked to regular changes in environmental conditions.

[0088] In particular, the machine learning model receives as input at least one matrix comprising measurement data, their timestamp data, and their associated supplementary data. This data includes deformation values ​​and their associated timestamps, as well as environmental values ​​and their associated timestamps.

[0089] The term "matrix" refers to a two-dimensional ordered data structure that groups several numerical or categorical values ​​representing different characteristics or measurements associated with a given instant or sequence of instants. This matrix serves as input to the machine learning model, encapsulating all the relevant information to generate a time series describing the expected deformation changes. For example, an input matrix might contain the different time measurements in rows and the current deformation value, the last N deformation values, the ambient temperature, the light intensity, the time of day, and the day of the year in columns. It might also include derived variables, such as the temperature change over the last few hours or the rate of change of the deformation.In some cases, the matrix can be enriched with contextual information, such as the sensor ID or codes indicating specific events (maintenance, silo filling, etc.). Furthermore, the matrix can be normalized or transformed to facilitate processing by the machine learning model, for example by applying min-max scaling or standardizing variables by column.

[0090] Next, the first processor 142 sequentially or in real time, for each strain sensor 110, uses the time series generated by its specific machine learning model to isolate the different strain components, based on the corresponding timestamps. Real-time processing allows for immediate data analysis, while sequential processing may be preferred for in-depth analysis or in situations where computing resources need to be optimized.

[0091] The expression "real-time" means that data processing and the isolation of deformation components are performed instantaneously or almost instantaneously, with a negligible delay compared to the acquisition of measurements. This notion of real-time implies that the system 100 is capable of providing processed and usable results at a sufficiently high rate to allow continuous and reactive monitoring of the structure's condition. By way of illustration, real-time processing can mean that component isolation is performed and results updated every second, allowing for the rapid detection of any anomaly in the silo's behavior. It can also refer to a system capable of processing a continuous stream of data from multiple sensors, in now has a latency of less than 100 milliseconds between receiving a measurement and producing its processed value. In some cases, real-time processing may involve the use of optimization techniques such as batch processing or sliding windows to efficiently handle large volumes of data. Furthermore, the concept of real-time can be extended to the system's ability to trigger alerts or automatic actions in response to detected critical situations without any perceptible delay.

[0092] - in the remote server: the first storage memory

[0093] In the invention, the first storage memory 143 is specifically designed to store, for each strain sensor 110, its specific previously trained machine learning model.

[0094] In addition, the first storage memory 143 stores the time series describing the deformation variations generated by the machine learning model specific to each deformation sensor 110.

[0095] The term "storage memory" refers to a device or physical medium capable of recording and retaining digital data permanently for later use by the processor 142 or other system components 100. This memory is designed to store both the raw data collected by the sensors, the machine learning models, and the results of the processing performed. For example, the storage memory could be a solid-state drive (SSD) integrated into the remote server, offering high read and write performance for real-time data processing. It could also take the form of a distributed storage system, spread across several nodes in a cluster, to ensure redundancy and scalability. In some cases, the storage memory could include long-term media such as magnetic tapes or optical discs for archiving historical data.Furthermore, the term can encompass hierarchical storage solutions, combining different technologies to optimize the trade-off between performance, capacity, and cost.

[0096] First embodiment of the first aspect of the invention: temporal categorization of environmental data for the isolation of deformation components

[0097] In a first embodiment of the first aspect of the invention, the first processor 142 applies time series describing the deformation variations generated by the machine learning model according to the following steps.

[0098] First, the first processor 142 categorizes the local environmental measurement data extracted from the second measurement signal 121 into predetermined time periods corresponding to predetermined ranges of variation.

[0099] The term "categorize" refers to the action of grouping elements with similar characteristics into predefined classes or categories, according to specific criteria. This operation aims to structure and simplify the analysis of a dataset by organizing it in a logical and consistent manner. As an example, the first processor 142 categorizes local environmental measurement data by distributing it into different time periods corresponding to predetermined ranges of variation. This might involve classifying temperature measurements according to whether they fall within value ranges representative of the morning, afternoon, or night. More specifically, this data can be classified according to predetermined 6-hour time periods, corresponding respectively to the typical variation ranges of the morning (6 a.m. to 12 p.m.), afternoon (12 p.m. to 6 p.m.), evening (6 p.m. to midnight), and night (midnight to 6 a.m.). Similarly, sunshine data can be categorized according to light intensity thresholds corresponding to clear, partly cloudy, or overcast conditions.Alternatively, these measurements can be categorized according to monthly periods, associated with variation ranges representative of the seasons: spring (March to May), summer (June to August), autumn (September to November), and winter (December to February). Furthermore, categorization can be applied to the deformation variations themselves, grouping them according to their amplitude or temporal dynamics. In this case, deformation data can be analyzed over 24-hour periods, with variation ranges reflecting the loading and unloading cycles of the silo.

[0100] Next, the first processor 142 determines categorization thresholds.

[0101] The term "categorization thresholds" refers to limit values ​​used to delimit the different classes or categories into which measurement data are distributed. These thresholds are determined in such a way as to optimize the separation between groups and to facilitate the interpretation of results, based on relevant statistical or physical criteria. For example, temperature categorization thresholds can be set according to the quartiles of the statistical distribution of measurements, making it possible to obtain different classes: low temperatures (below the first quartile), medium temperatures (between the first and third quartiles), high temperatures (above the third quartile), and extreme temperatures (above the 5th and 95th percentiles).Similarly, solar irradiance categorization thresholds can be defined in terms of energy irradiance levels, expressed in W / m²: low (less than 200 W / m²), moderate (between 200 and 600 W / m²), high (between 600 and 1000 W / m²), and very high (more than 1000 W / m²). Furthermore, deformation categorization thresholds can be established based on elastic deformation limits. plastic of the material constituting the silo, or stress levels corresponding to different safety factors.

[0102] In practice, the categorization thresholds can be determined based on the statistical distribution of local environmental measurement data in the training dataset.

[0103] The expression "statistical distribution of local environmental measurement data" refers to the distribution of measured values ​​for a given environmental quantity, as it can be characterized by various statistical parameters and indicators. This distribution reflects the variability and dispersion of the measurements around their central tendency, as well as the possible presence of outliers or extreme values. By way of illustration, the statistical distribution of ambient temperatures can be described by its mean, median, and standard deviation, which respectively indicate the central value, midpoint, and overall dispersion of the measurements. It can also be represented graphically by a histogram or a probability density curve, highlighting the modes (peaks) and tails of the distribution.Similarly, the distribution of sunshine measurements can be characterized by its skewness and kurtosis coefficients, which quantify the degree of asymmetry and concentration of values ​​around the mean, respectively. Furthermore, statistical goodness-of-fit tests (such as the Kolmogorov-Smirnov test) can be applied to determine whether the empirical distribution of the data corresponds to a known theoretical distribution (normal, log-normal, Weibull, etc.).

[0104] In an approach complementary to statistical analysis, these categorization thresholds can also be determined based on transient environmental phenomena affecting the support foot of silo 11.

[0105] The expression "transient environmental phenomena affecting the silo support base" refers to temporary variations in local ambient conditions that can influence the mechanical behavior and deformation measurements of the silo support base 11, without reflecting an intrinsic change in its structural state. These phenomena are generally linked to seasonal periodic cycles, which repeat, for example, every 24 hours. For example, temperature fluctuations during the day, induced by alternating periods of sunlight and nighttime cooling, can cause thermal expansion and contraction of the metal support base, superimposed on the mechanical deformations due to the weight of the stored materials.Similarly, variations in ambient relative humidity, accentuated during periods of rain or fog, can alter the frictional properties of contacting surfaces and affect the transmission of forces within the structure. Furthermore, [these factors]... Transient phenomena such as gusts of wind or thunderstorms can generate vibrations or point shocks on the support foot of silo 11, temporarily disrupting deformation measurements without compromising its long-term integrity.

[0106] For each predetermined time period, the first processor 142 performs several operations.

[0107] First, it determines at least one statistical value representative of the deformation components related to environmental phenomena, extracted from local environmental measurement data from the second measurement signals with their associated timestamp data for the period.

[0108] The expression "representative statistical value of the deformation components for the period" designates a single numerical indicator that summarizes and characterizes the distribution of deformation components related to environmental phenomena for a given time period, as determined by the environmental data categorization process. This representative value makes it possible to summarize, in a concise and relevant manner, all the deformation components to be identified in the measurements for the period under consideration, taking into account their variability and central tendency. By way of illustration, the representative statistical value may be the arithmetic mean of the environmental components, which reflects their overall level and mutually compensates for positive and negative deviations.It can also correspond to the median of the components, which is less sensitive to extreme values ​​and provides a robust indication of their central tendency. In some cases, the representative value may be a specific quantile (for example, the 95th percentile) to focus on the most important components. Furthermore, other indicators such as the standard deviation or the interquartile range can be used in conjunction to characterize their dispersion and their typical range of variation over the period.

[0109] Next, the first processor 142 stores, in the first storage memory 143, the statistical value in association with the local environmental measurement data and their timestamps.

[0110] Finally, the first processor 142 generates deformation components specific to the position of the deformation sensor 110 on the support foot of silo 11 using the statistical value representative of the deformation components for the corresponding time period.

[0111] Thus, the first processor 142 produces adapted and optimized strain component values ​​for each strain sensor using separate and independent machine learning models. This individualization approach allows each model to accurately capture the specific characteristics related to the sensor's specific location on the silo 11 support structure and the unique local environmental conditions it is subjected to. For example, for a sensor located on the south face of the silo 11 support, which, in the Northern Hemisphere, is directly exposed to solar radiation, the generated deformation components will incorporate the predominant influence of temperature and solar irradiance variations, based on representative statistical values ​​determined for daytime periods. Conversely, for a sensor positioned on the north face, which, in the Northern Hemisphere, is subject to shading conditions, the deformation components will be based primarily on representative nighttime or winter values, which are less impacted by heating and cooling cycles.Similarly, the deformation components applied to sensors near the base of the silo 11 support foot, potentially affected by soil moisture, will differ from those of sensors in the upper part, which are more sensitive to wind effects. Furthermore, the generation of specific deformation components may involve additional weighting coefficients or correction terms to account for local influencing factors such as the geometry of the silo 11 support foot, the orientation of the instrumented faces, or the proximity of heat sources.

[0112] A particular embodiment of the first embodiment of the first aspect of the invention: categorization of the ranges of variation of environmental data for the isolation of deformation components

[0113] In a particular embodiment of the first embodiment of the first aspect of the invention, the predetermined ranges of variation comprise three distinct categories.

[0114] The first category is a range of small variations, below a first predetermined threshold.

[0115] In practice, the range of small variations corresponds to the smallest variations relative to the central tendency or the reference value of the quantity under consideration. For example, for ambient temperature measurements, the range of small variations may include all values ​​less than 5°C relative to the seasonal average. Similarly, for solar radiation data, this range may include irradiance levels below 200 W / m², corresponding to low-light conditions. Furthermore, in the case of deformation measurements of a silo support base 11, the range of small variations may include deformation values ​​less than 10% of the maximum permissible deformation, reflecting a state of reduced mechanical stress.

[0116] Next, the second category is a range of median variations, between the first threshold and a second predetermined threshold.

[0117] In practice, the median range of variation corresponds to intermediate amplitude variations located on either side of the central tendency or the reference value of the quantity under consideration. By way of illustration, for ambient temperature measurements, the median range of variation may include all values ​​between 5°C and 10°C of deviation from the seasonal average. Similarly, for solar radiation data, this range may include energy irradiance levels between 200 W / m² and 600 W / m², corresponding to moderate light conditions. Furthermore, in the case of deformation measurements of a silo support base 11, the median range of variation may include deformation values ​​between 10% and 50% of the maximum permissible deformation, reflecting an average state of mechanical stress.

[0118] Finally, the third category is a range of high variations, above the second predetermined threshold.

[0119] In practice, the range of variation corresponds to the largest variations relative to the central tendency or the reference value of the quantity under consideration. For example, for ambient temperature measurements, the range of high variation may include all values ​​greater than 10°C relative to the seasonal average. Similarly, for solar irradiance data, this range may include levels of irradiance greater than 600 W / m², corresponding to conditions of high luminosity. Furthermore, in the case of deformation measurements of a silo support base 11, the range of high variation may include deformation values ​​greater than 50% of the maximum permissible deformation, reflecting a state of significant mechanical stress, potentially critical to the integrity of the structure.

[0120] It is important to note that the first and second thresholds are determined as categorization thresholds for these predetermined ranges of variation.

[0121] Second embodiment of the first aspect of the invention: sequential or real-time processing of measurements by contextual analysis of environmental variations

[0122] The first processor 142 applies sequentially or in real time the process of isolating the deformation components to the deformation measurements according to the following steps.

[0123] First, for each new deformation measurement received at a given time via the first measurement signal 111 with its associated timestamp data, the first processor 142 performs several operations.

[0124] First, the first processor 142 determines a range of variation of the local environmental measurement data over a predetermined time period including the last measurement instant.

[0125] In other words, the first processor 142 identifies the extent of fluctuations in locally measured environmental parameters over a defined time interval that encompasses the precise moment of the new deformation measurement. This operation aims to contextualize the deformation measurement within its immediate environment. For example, the processor 142 can determine that the temperature varied between 15°C and 25°C over a period of 2 hours preceding the time of measurement. Similarly, it can establish that the solar irradiance fluctuated between 200 W / m² and 800 W / m² over the same period. Furthermore, the processor 142 can identify a range of relative humidity variation between 40% and 60% during the interval considered.

[0126] Secondly, the first processor 142 selects, from among the time series generated by the specific machine learning model, a time series corresponding to the local environmental measurement data extracted from the second measurement signals with their associated timestamp data for the range of variation.

[0127] In other words, the first processor 142 selects the most appropriate sequence of deformation components based on the observed environmental conditions. This selection relies on the correspondence between the determined ranges of variation and the time series previously generated by the machine learning model. For example, the processor 142 can select a time series of deformation components specific to sunny summer days when the range of temperature and sunshine variation corresponds to these conditions. Similarly, it can select a time series suitable for winter night periods if the environmental data indicate low temperatures and a lack of sunshine. Furthermore, the processor 142 can select an intermediate time series for mid-season or variable weather conditions.

[0128] Next, the first processor 142 extracts from the time series a deformation component corresponding to two elements.

[0129] The first element is the time of the new deformation measurement determined from the timestamp data associated with the first measurement signal 111.

[0130] By way of example, the instant can be expressed as a precise date and time, such as "2023-11-04 15:32:47.123". It can also be represented as a number of seconds elapsed since a fixed time reference, such as 1 January 1970 at 00:00:00 UTC. In addition, the instant can include time zone information or be converted to Coordinated Universal Time (UTC) to facilitate synchronization with other data.

[0131] The second element consists of local environmental measurement data extracted from the second measurement signals with their associated timestamp data for the determined range of variation.

[0132] By way of illustration, this data may include the average, minimum, and maximum temperatures observed during the range of variation, as well as the temperature trend curve over this period. Similarly, it may include the average light intensity and its variation profile over the interval considered. Furthermore, this data may encompass other relevant parameters such as relative humidity, wind speed and direction, or atmospheric pressure, all extracted for the specific range of variation.

[0133] Finally, the first processor 142 applies the strain component to the strain measurement to obtain an isolated measurement.

[0134] The term "isolated measurement" refers to the result of the process of decomposing a total deformation measurement into its various constituent components. This isolated measurement specifically represents the deformation component related to the weight of the contents, obtained after the identification and quantitative separation of the other components (thermal, solar, etc.). For example, an isolated measurement of 0.5 mm represents the deformation solely due to the weight of the contents, once the components related to thermal expansion (e.g., 0.05 mm) and the effect of solar radiation (e.g., 0.03 mm) have been identified and separated from the initial total measurement of 0.58 mm. This isolated measurement thus makes it possible to accurately assess the actual load borne by the support foot, independently of transient environmental effects.

[0135] Thus, biases introduced by transient environmental phenomena are eliminated / reduced in order to isolate, in the total deformation measurement of the silo 10 support base, the component related solely to the weight of the contents. The measured total deformation is in fact composed of several components: that related to the weight of the contents, that related to temperature, that related to sunlight, and that related to other undetermined phenomena. For example, the processor 142 can subtract a deformation component of 0.05 mm from the raw deformation measurement to isolate the thermal expansion of the material due to a temperature increase. Similarly, it can apply a multiplicative factor of 0.98 to the measurement to isolate the effect of strong solar exposure on the sensor.In addition, the 142 processor can combine several deformation components, such as adding a 0.02 mm offset to isolate the effect of humidity, followed by normalization based on ambient temperature.

[0136] Third embodiment of the first aspect of the invention: generation of time series by analysis of historical data and exogenous environmental variables

[0137] In a third embodiment according to the first aspect of the invention, the machine learning model performs certain operations during its use.

[0138] First, the machine learning model receives as input at least one matrix comprising two types of data.

[0139] The first type of data includes previous strain values ​​and their timestamps associated with measurement signals.

[0140] By way of illustration, this first type of data may include the last 100 strain values ​​recorded by the sensor, spaced one second apart, with their respective timestamps in the format "YYYY-MM-DD HH:MM:SS". It may also consist of a sample of measurements taken at regular intervals over a given period, such as the average hourly strain values ​​over the last 24 hours. Furthermore, this data may be preprocessed to extract relevant features, such as strain variations between two successive times or long-term trends.

[0141] Next, the second type of data includes at least one exogenous variable corresponding to the local environmental measurement data extracted from the second measurement signals with their associated timestamp data and complementary data.

[0142] By way of example, an exogenous variable could be the average ambient temperature over the last 10 minutes, updated with each new model prediction. Another exogenous variable could be the maximum solar irradiance observed during the last hour, expressed in watts per square meter. In addition, further exogenous variables could be included, such as relative humidity, average wind speed, or atmospheric pressure, to provide a more complete environmental context for the model.

[0143] The term "exogenous variable" refers to a measurable quantity that influences the behavior of a system without being directly related to its internal parameters. In the specific context of the machine learning model, exogenous variables are local environmental data that affect the deformation measurements of the silo 10 support foot, without depending on the deformation state itself. For example, ambient temperature is an exogenous variable because it can cause thermal expansion or contraction of the silo 10 support foot, thus changing the measured deformation values, independently of the actual load supported. Similarly, sunlight is an This is an exogenous variable, as it can influence the surface temperature of the silo 11 support base and, consequently, the deformation measurements, without being directly related to the mechanical stress state of the structure. Furthermore, meteorological variables such as wind speed or air humidity can also constitute relevant exogenous variables, modifying the operating conditions of the deformation sensor or inducing second-order effects on the structure.

[0144] It is important to note that the exogenous variable makes it possible to identify the different components of the total measured deformation.

[0145] Indeed, by integrating exogenous variables into the machine learning model, it becomes possible to decompose the total measured deformation by isolating the components due to transient environmental phenomena from the component related to the mechanical stresses induced by the weight of the contents in the silo 11 support base. For example, a sudden increase in ambient temperature can be detected by the corresponding exogenous variable and interpreted as a transient phenomenon, thus allowing the model to correct the associated deformation measurements accordingly. Similarly, an abnormal variation in sunlight can be identified via the dedicated exogenous variable and signal a one-off meteorological event, such as a passing cloud, which does not compromise the structural integrity of the silo 11 support base, but which temporarily affects the measurements.Furthermore, the correlation between several exogenous variables can provide a more reliable indication of the occurrence of a transient environmental phenomenon, such as a storm combining wind gusts, a drop in atmospheric pressure, and intense rainfall.

[0146] Finally, the machine learning model generates a time series describing the expected deformation variations under the environmental conditions corresponding to the received input matrix.

[0147] Fourth embodiment of the first aspect of the invention: multi-scale analysis of deformation components by differentiated treatment of transient environmental phenomena

[0148] In a fourth embodiment according to the first aspect of the invention, the transient environmental phenomenon detection sensor 120 and the machine learning model perform certain operations during the use of the machine learning model.

[0149] First, the transient environmental phenomenon detection sensor 120 generates at least two distinct exogenous variables.

[0150] The term “distinct” refers to the differentiated and independent nature of the exogenous variables generated by the transient environmental phenomenon detection sensor 120. These variables are considered distinct, Because they represent different aspects of local environmental conditions and are processed separately by the machine learning model. For example, one distinct exogenous variable might be the instantaneous ambient temperature, while a second distinct exogenous variable might be the average light intensity over a given period. Similarly, one distinct exogenous variable might represent the relative humidity of the air, while another might correspond to the wind speed. Furthermore, the exogenous variables might be distinct in their temporal nature, one reflecting short-term variations and the other long-term trends in environmental conditions.

[0151] In practice, the first exogenous variable corresponds to first local environmental measurement data extracted from the second measurement signal 121 with its associated timestamp data.

[0152] Then, the second exogenous variable corresponds to second local environmental measurement data extracted from the second measurement signal 121 with its associated timestamp data, and which are characteristic of seasonal cycles.

[0153] The expression "seasonal cycles" refers to periodic variations in environmental conditions that recur with a constant frequency. In the context of machine learning, a cycle refers to a variation that recurs regularly over time, such as 24-hour temperature variations (daily cycle) or changes in sunlight that repeat daily.

[0154] Thus, the data seconds reflect the periodic variations in environmental conditions that occur over defined time scales. This data captures recurring patterns that can influence the behavior of the silo 11 support foot of silo 10. By way of illustration, the data seconds may include the average temperature, which varies according to a 24-hour cycle with higher values ​​during the day and lower values ​​at night.

[0155] Then, the machine learning model receives as input at least one matrix comprising the first data and the second data generated by the transient environmental phenomenon detection sensor 120.

[0156] Thus, the machine learning model uses a structured data format (the matrix) that combines short-term (first data) and long-term (second data) information to generate sequences of deformation components specific to each time scale. For example, the input matrix might contain the instantaneous temperature, relative humidity, and current sunshine (first data), as well as the average temperature of the last seven days and the length of the day (second data).

[0157] Next, the machine learning model generates time series that describe the deformation variations corresponding to the data contained in the input matrix.

[0158] Indeed, the machine learning model preferentially generates a single time series that describes the shape of the deformation variations related to environmental phenomena corresponding to the deformation values, environmental data, and their respective timestamps contained in the input matrix. This time series, based on learning the characteristic deformation patterns observed for similar combinations of this input data, makes it possible to identify and isolate the deformation component related to environmental phenomena. Alternatively, the model can also generate several distinct time series, for example, to separately describe the variations related to short-term thermal effects and long-term seasonal effects identified from the input matrix data.

[0159] Second aspect of the invention: a training system for a machine learning model for isolating the deformation components of a silo support foot

[0160] The second aspect of the invention relates to a system 200 for training and applying a machine learning model for processing deformation measurements of a silo support foot 11. This system 200 integrates two complementary and interdependent processes: a model training process and a process for processing the measurements sequentially or in real time. These two processes, although distinct in their function, form a coherent whole to ensure the efficiency and accuracy of the monitoring system.

[0161] The continuous interaction between the training and processing phases, whether sequential or in real time, allows for constant improvement in system performance. New data collected during sequential or real-time processing are regularly integrated into the training dataset, thus enabling progressive refinement of the models. This feedback loop ensures that the system remains adapted to the specific site conditions, even as these conditions change over time.

[0162] As illustrated in [Fig. 3], a second aspect of the invention relates to a training system 200 for one or more individualized machine learning models, each machine learning model for processing deformation measurements from sensors installed on the support feet of silo 11 of a silo 10 for storing or mixing at least one bulk material. The system can use either a single model to process data from several sensors having similar characteristics or installation conditions, or Customized models specifically designed for each sensor according to needs. This flexibility allows for optimized data processing based on the installation's characteristics, environmental conditions, and monitoring objectives.

[0163] Thus, System 200 aims to develop and optimize an algorithm capable of automatically adjusting deformation measurements to improve the accuracy of silo content estimation. This System 200 uses artificial intelligence techniques to learn from historical data and adapt to the specific conditions of each installation. For example, this System 200 may include a powerful processor 210 that analyzes thousands of deformation measurements collected over several months, correlating them with variations in temperature, sunlight, and other environmental factors. It may also include sophisticated algorithms that identify recurring patterns in the deformations of the silo support base 11, such as diurnal or seasonal cycles, to distinguish them from variations due to silo filling or emptying.In addition, this System 200 can incorporate cross-validation mechanisms that test the model's performance on data not used during training, thus ensuring its ability to generalize correctly.

[0164] In practice, the system 200 includes at least a second processor 210 and at least a second storage memory 220 of the same type as the first processor 142 and the first storage memory 143.

[0165] - the second processor

[0166] In the invention, the second processor 210 is specifically designed to perform several tasks.

[0167] First, the second processor 210 identifies time periods during which the deformation variations are mainly due to transient environmental phenomena. To do this, the second processor 210 analyzes the deformation variations from measurements acquired by at least one deformation sensor 110, as explained above.

[0168] In other words, it involves isolating specific time intervals where the measured changes in the structure of the silo support foot 11 are primarily attributable to temporary external factors rather than to changes in the silo contents. This identification relies on a thorough analysis of the data collected by the deformation sensor 110. For example, the processor 210 can identify nighttime periods when the silo is not in operation, but where the deformation measurements fluctuate in correlation with variations in ambient temperature. It can also identify sunny days where the deformations follow a cyclical pattern corresponding to the solar exposure of the silo support foot 11, regardless of filling or emptying operations. Furthermore, the processor 210 can detect episodes of strong wind or heavy rainfall which induce transient disturbances in deformation measurements, without relation to the actual level of filling of the silo.

[0169] Next, the second processor 210 detects regular variation patterns corresponding to measurements acquired by at least one transient environmental phenomenon detection sensor 120, as explained above.

[0170] Here, the aim is to identify recurring patterns in environmental data that coincide with certain observed deformation variations. This detection makes it possible to establish correlations between ambient conditions and the mechanical behavior of the silo 11 support base. For example, the processor 210 can recognize a daily pattern in temperature measurements that is reflected in the cyclical variations in the deformation of the silo 11 support base, with maximum expansion in the middle of the day and contraction during the night. It can also identify weekly patterns related to surrounding industrial activities, such as an increase in ground vibrations every Monday morning due to the restart of machines in a neighboring factory.In addition, the 210 processor can detect seasonal patterns in sunlight data that correspond to gradual changes in the long-term behavior of the silo 11 support foot, such as a slow drift in deformation measurements over months.

[0171] After this, the second processor 210 constitutes a training dataset.

[0172] The term "training dataset" refers to a structured set of carefully selected and prepared historical information, which serves as the basis for training the machine learning model. This dataset contains both the deformation measurements and the associated environmental data, organized in such a way as to allow the model to learn the complex relationships between these variables. By way of illustration, a training dataset might include a full year of deformation measurements taken every minute, along with the corresponding temperature, sunlight, and humidity readings. It might also include specific periods when the silo underwent known changes in fill level, thus serving as a reference for calibrating the model.In addition, the dataset may contain information on exceptional events, such as storms or heat waves, allowing the model to learn to handle extreme situations.

[0173] To do this, the second processor 210 extracts the deformation data and their timestamps from the measurements acquired by the deformation sensor 110.

[0174] Then, the second processor 210 centers the deformation data to eliminate the offsets between time periods.

[0175] The term “centering” refers to a mathematical operation applied to deformation data that involves subtracting a reference value (usually the mean) from each measurement to eliminate systematic shifts between different time periods. This operation makes it easier to compare deformation variations regardless of their absolute magnitude. For example, the 210 processor can center deformation data by calculating the average over a full day and subtracting this value from each individual measurement, thus highlighting relative fluctuations rather than absolute values. It can also apply sliding-window centering, where the reference value is recalculated for each fixed time interval, thereby accommodating slow drifts in the signal.Furthermore, centering can be performed separately for different silo filling ranges, allowing comparison of deformation behaviors under similar load conditions.

[0176] Finally, the second processor 210 concatenates the data while preserving their temporal consistency.

[0177] The term "concatenate" refers to the action of combining several time series data into a single continuous sequence, while preserving the chronological order and consistency of the information. This operation makes it possible to create a unified training dataset from multiple sources or acquisition periods. By way of illustration, the 210 processor can concatenate deformation measurements from several sensors installed on different support legs of the same silo, aligning them precisely in time thanks to their timestamps. It can also assemble discontinuous environmental data series, such as meteorological readings interrupted by maintenance periods, by inserting special markers to indicate the interruptions.Furthermore, concatenation may involve merging data at different sampling frequencies, requiring interpolation or subsampling to ensure overall temporal consistency.

[0178] In the invention, the second processor 210 then trains one or more machine learning models to process the data from the strain sensors. Depending on the needs and characteristics of the installation, the system can use either a single model for a group of sensors with similar conditions, or separate models for sensors requiring specific processing. This flexible approach makes it possible to optimize the processing according to the particular exposure and mechanical stress conditions specific to each sensor position or group of sensors. As a particularly advantageous example, the machine learning model can be a Long Short-Term Memory (LSTM) neural network. This architecture is particularly The LSTM network is well-suited for processing deformation time series due to its ability to capture long-term dependencies in the data via its dedicated memory cell. It can be configured with two successive layers: the first comprising 128 units and the second 256 units, followed by a dense cone layer. This configuration allows for high accuracy in reconstructing missing data. The LSTM model proves particularly effective for processing structural monitoring data, especially during extreme weather events, thanks to its ability to simultaneously analyze temporal and spatial correlations between different measurements.

[0179] To this end, the second processor 210 prepares at least one input matrix comprising data extracted from the measurement signals with their associated timestamp data and complementary data. This data includes the strain values ​​and their associated timestamps, as well as the environmental values ​​and their associated timestamps.

[0180] Here, it is an iterative process by which the system 200 optimizes the model's internal parameters by repeatedly comparing them with historical data. The model learns to identify correlations between environmental variables and the deformation signal, allowing the extraction of a time series representative of the deformations induced by environmental phenomena. This time series makes it possible to isolate the deformation components to be applied to the deformation measurements in order to isolate the component related to the weight of the contents. As an example, the processor 210 can employ reinforcement learning methods, where the model is rewarded for accurate predictions and penalized for errors, thus encouraging it to refine its estimates over time.In addition, training may include cross-validation phases, where the dataset is divided into subsets to test the model's ability to generalize to unseen data.

[0181] Finally, the second processor 210 uses the input matrix to train the model to generate predictions of the deformation components adapted to the specific position of each sensor on the silo support foot 11. These deformation components take into account the sensor's exposure to local environmental conditions and its specific mechanical response according to its location on the structure.

[0182] Here, the process involves the system integrating multidimensional information into the model learning, simultaneously considering the influence of environmental conditions and the specific mechanical characteristics of each installation. The input matrix, which contains both deformation measurements and environmental data with their respective timestamps, allows the model to establish complex correlations between these different variables. As an illustration, the 210 processor can train the model to recognize how the thermal expansion of the silo 11 support base varies not only with ambient temperature but also with the silo's orientation relative to the sun, using the solar radiation data included in the input matrix. It can also learn to distinguish deformations due to temperature variations from those caused by changes in fill level by analyzing the temporal patterns of deformation measurements in relation to daily thermal cycles. Furthermore, the model can be trained to take into account the specific characteristics of each sensor, such as its exact position on the silo 11 support base or its individual sensitivity to mechanical stresses, using unique identifiers included in the input matrix for each measurement point.

[0183] - the second storage memory

[0184] Finally, the second storage memory 220 is specifically designed to store three elements: the training dataset, the machine learning model, and the time series generated by the machine learning model.

[0185] First embodiment of the second aspect of the invention: multi-approach detection of the periods of influence of environmental phenomena on the deformation components

[0186] In a first embodiment of the second aspect of the invention, the second processor 210 identifies the time periods according to at least one of the following three approaches.

[0187] The first approach consists of detecting periods of silo inactivity, namely, identifying time intervals during which the silo does not undergo significant changes in its contents, whether through filling or emptying. This approach aims to isolate the times when the measured deformation variations are primarily due to environmental factors rather than storage operations. For example, detecting periods of inactivity could include identifying nighttime hours in a cement plant, where loading and unloading operations are generally suspended between 10 p.m. and 6 a.m. It could also involve recognizing public holidays in a food processing plant, where production activity, and therefore the handling of raw materials stored in silos, is interrupted.Furthermore, this approach can be applied to the detection of planned maintenance periods, such as during an annual one-week shutdown in a concrete plant, during which the cement and aggregate silos remain at a constant level.

[0188] Next, the second approach involves the detection of regular and cyclical variation patterns characteristic of transient environmental phenomena, at Namely, the identification of recurring patterns in deformation measurements that correspond to predictable changes in environmental conditions. This approach makes it possible to distinguish deformation variations due to cyclical external factors from those caused by changes in the silo's contents. For example, this detection could involve identifying a daily cycle of thermal expansion and contraction in a metal grain storage silo, with maximum expansion around 2 p.m. when sunlight is strongest, and minimum contraction around 5 a.m. It can also be applied to recognizing a weekly pattern in the deformation measurements of a cement silo, where vibrations induced by nearby road traffic are greater from Monday to Friday than during the weekend.Furthermore, this approach can help identify seasonal variations, such as the gradual changes in the average deformation of a de-icing salt silo between summer and winter, due to differences in ambient temperature and humidity.

[0189] Finally, the third approach involves analyzing strain variations correlated with local environmental measurement data, namely, studying the relationships between observed changes in silo strain measurements and simultaneous fluctuations in environmental conditions. This approach aims to establish causal links between external factors and the silo's mechanical behavior, thereby enabling better isolation and correction of the effects of transient environmental phenomena. As an example, this analysis could include studying the correlation between strain peaks in a chemical storage silo and episodes of strong winds, by comparing strain sensor data with those from a local anemometer.It can also involve examining the relationship between variations in the deformation of a biomass storage silo and changes in relative humidity, using measurements from a nearby hygrometer. Furthermore, this approach can be applied to analyzing the impact of rainfall on the deformation of a salt storage silo, by correlating deformation data with readings from a rain gauge located on the site.

[0190] It is important to note that the time periods are defined according to the cyclic phenomenon being studied. In a particular implementation, where daily seasonal phenomena are analyzed, these periods include periods of at least 24 consecutive hours during which the silo exhibits regular variations in the measurements of the strain sensor 110 corresponding to daily cycles. However, the analysis of cyclic phenomena of different frequencies (shorter or longer) would require adjusting the duration of these time periods accordingly.

[0191] Indeed, observing phenomena over at least one complete diurnal cycle makes it possible to capture all daily environmental variations. For example, a time period could correspond to a full week of observation of a grain storage silo during the harvest season, where deformation measurements show a repetitive pattern with maximum expansion in the mid-afternoon and maximum contraction just before dawn, despite filling and emptying operations. It could also refer to a 72-hour period of continuous monitoring of a cement silo over a long weekend, revealing deformation cycles linked to variations in temperature and sunlight, in the absence of industrial activity.Furthermore, this approach can be applied to the analysis of a full month of deformation data from a de-icing salt storage silo during the summer, highlighting stable daily cycles of thermal expansion and contraction, without disturbance due to the use of the contents.

[0192] Second embodiment of the second aspect of the invention: categorization and statistical processing of environmental data for learning deformation components

[0193] In a second embodiment of the second aspect of the invention, the second processor 210 performs several tasks.

[0194] First, as explained above with regard to the first processor 142, the second processor 210 categorizes the local environmental measurement data extracted from the second measurement signal 121 into predetermined time periods corresponding to predetermined ranges of variation.

[0195] Next, as explained above regarding the first processor 142, the second processor 210 determines categorization thresholds based on two elements. The first element is the statistical distribution of local environmental measurement data in the training dataset. The second element concerns transient environmental phenomena affecting the silo support foot 11.

[0196] For each predetermined time period, as explained above with regard to the first processor 142, the second processor 210 performs two operations.

[0197] First, it determines at least one statistically representative value of the deformation components for the period.

[0198] Secondly, the second processor 210 stores, in the second storage memory 220, the statistical value representative of the deformation components for the corresponding time period in association with the local environmental measurement data and their timestamps.

[0199] A particular embodiment of the second embodiment of the second aspect of the invention: categorization of the ranges of variation of environmental data for learning deformation components

[0200] In a particular embodiment of the second embodiment of the second aspect of the invention, the predetermined ranges of variation comprise three distinct categories.

[0201] In practice, as explained above concerning the particular embodiment of the first embodiment of the first aspect of the invention, the first category is a range of small variations, below a first predetermined threshold. Next, the second category is a range of intermediate variations, between the first threshold and a second predetermined threshold. Finally, the third category is a range of large variations, above the second predetermined threshold.

[0202] It is important to note that the first and second thresholds are determined as categorization thresholds for these predetermined ranges of variation.

[0203] Third embodiment of the second aspect of the invention: adaptive updating of the model by detecting significant changes in environmental components

[0204] In a third embodiment of the second aspect of the invention, the second processor 210 periodically updates the machine learning model according to the following steps.

[0205] The term “update” refers to the process by which the second processor 210 periodically refreshes and improves the machine learning model to maintain its accuracy and relevance in the face of changing environmental conditions. This operation aims to adapt the model to significant changes in local environmental measurement data that may occur over time. For example, updating may involve adjusting model parameters to account for seasonal temperature variations that affect silo deformation. It may also involve incorporating new training data reflecting exceptional weather conditions, such as a prolonged heat wave, that were not represented in the initial dataset.In addition, the update may include revising the anomaly detection thresholds based on the progressive wear of the silo materials, thus modifying its response to mechanical and environmental stresses.

[0206] To do this, the second processor 210 detects a significant change in local environmental measurement data over a predetermined period corresponding to a seasonal variation.

[0207] The expression "significant change" refers to a substantial and statistically relevant variation in local environmental measurement data, which This justifies a reassessment and adaptation of the machine learning model. A change is considered significant when it exceeds a predetermined threshold and is likely to affect the accuracy of the model's predictions. For example, a significant change could be a 5°C increase in the average daily temperature over a one-month period, indicating a change of season that necessitates a model update. It could also be a change in wind patterns, such as a 50% increase in the frequency of gusts exceeding 80 km / h, which could influence the mechanical stresses experienced by the silo. Furthermore, a significant change can be detected in the case of a change in the average humidity level, for example, increasing from 60% to 80% over a prolonged period, which could affect the properties of the stored materials and, consequently, the deformation of the silo.

[0208] The term “seasonal” characterizes the time scales over which environmental variations are analyzed to detect significant changes requiring model updates. This approach allows for consideration of both daily fluctuations and longer-term trends that can affect silo behavior. For example, a seasonal analysis might focus on day-night temperature cycles, which can induce repeated expansion and contraction of the silo over a 24-hour period. The seasonal analysis might also consider gradual changes in average temperature over several months, reflecting the transition from summer to winter.Furthermore, this approach can combine these two scales, examining, for example, how the amplitude of daily temperature variations evolves over the seasons, moving from large deviations in summer to more moderate fluctuations in winter.

[0209] This detection comprises two steps. The first step consists of calculating a difference between current local environmental measurement data and that used for training the machine learning model, taking into account the associated timestamps. The second step involves comparing the difference to a predetermined threshold.

[0210] By way of illustration, the first step could involve calculating the difference between the average temperature of the last 30 days and that of the same period used during the initial training of the model. The second step could then compare this difference to a threshold of 3°C, beyond which an update would be triggered. Similarly, for solar irradiance, the 200 system could calculate the difference between the cumulative solar energy over a month and the reference value, then compare this difference to a threshold of 15% variation. Furthermore, this approach can be applied to combinations of parameters, such as the calculation of a heat stress index integrating temperature and humidity, the deviation of which would be compared to a predefined threshold to decide whether an update is necessary.

[0211] When the deviation exceeds the predetermined threshold, the second processor 210 performs three operations.

[0212] First, it automatically creates a new training dataset comprising strain values ​​and their timestamps, as well as local environmental measurement data and their timestamps.

[0213] Secondly, the second processor 210 retrains the machine learning model with the new dataset.

[0214] Third, it generates new time series.

[0215] In practice, this process ensures that the machine learning model remains relevant and accurate in its predictions, despite significant changes in the environment. For example, if a significant change is detected in the temperature regime, the processor 210 could first create a new dataset including strain and temperature measurements from the last three months, with hourly resolution. Then, it could retrain the model using this new data, adjusting its internal parameters to better reflect the updated relationship between temperature and strain. Finally, the processor 210 would generate new time series, for example, by calculating correction coefficients to distinguish and eliminate the influence of temperature on the strain measurements for each hour of the day, taking into account the new temperature profile.Furthermore, this process could be applied to other environmental variables, such as humidity or wind, by adapting the dataset, model retraining, and deformation component generation to the specific characteristics of each parameter.

[0216] Fourth embodiment of the second aspect of the invention: generation of time series describing deformation variations with exogenous variables

[0217] In a fourth embodiment of the second aspect of the invention, the machine learning model is a time series prediction model specifically designed to perform certain operations, as explained above with regard to the third embodiment according to the first aspect of the invention.

[0218] First, the time series prediction model receives as input at least one matrix comprising two types of data. The first type of data includes strain values ​​and their timestamps extracted from the first measurement signals with their associated complementary data. Second, the second type of data includes at least one exogenous variable corresponding to the data of local environmental measurements extracted from the second measurement signals with their associated complementary data.

[0219] The time series prediction model generates time series that describe the expected deformation variations under specific environmental conditions according to a two-phase process.

[0220] In a first phase, the system creates and trains a separate model for each category of values ​​of the exogenous variable, thus enabling it to learn the characteristic deformation patterns under different environmental conditions. These models are then saved in a database.

[0221] In a second phase, the system uses these pre-trained models sequentially or in real time. For each new measurement, it identifies the category of the current exogenous variable and uses the corresponding model to generate a time series describing the shape of the deformation variations related to environmental phenomena under these specific conditions.

[0222] Fifth embodiment of the second aspect of the invention: multi-scale analysis of deformation components by processing environmental variables

[0223] In a fifth embodiment according to the first aspect of the invention, the transient environmental phenomenon detection sensor 120 and the machine learning model perform certain operations when using the machine learning model, as explained above with regard to the fourth embodiment according to the first aspect of the invention.

[0224] The transient environmental phenomenon detection sensor 120 generates at least two distinct exogenous variables. The first exogenous variable consists of initial local environmental measurement data extracted from the second measurement signal 121, along with its associated timestamp data. The second exogenous variable consists of further local environmental measurement data extracted from the second measurement signal 121, along with its timestamp data and associated supplementary data. It is important to note that this further data is characteristic of seasonal cycles.

[0225] The machine learning model receives as input at least one matrix comprising the first and second data generated by the transient environmental phenomenon detection sensor 120. Then, the machine learning model generates separate time series as a function of the input matrix comprising the first and second data.

[0226] Conclusion

[0227] We have described and illustrated the invention. However, the invention is not limited to the embodiments we have presented. Indeed, many Combinations of variants, alternatives, embodiments, and implementations can be considered without requiring substantial modifications to the invention. Thus, an expert in the field can deduce other variants, alternatives, embodiments, and implementations by reading the description and the accompanying figures, and taking into account the economic, ergonomic, and dimensional constraints to be respected.

[0228] For example, the strain sensor and the sensor for detecting transient environmental phenomena can be integrated into a single physical device. In this configuration, the same housing can contain both the strain measurement elements and the environmental sensors, sharing a common electronic circuit for signal processing and transmission. This integration optimizes the footprint on the support foot while ensuring perfect synchronization of strain measurements and environmental data.

[0229] Furthermore, when an expression uses the term "at least one", this means that the element or characteristic in question may be present in a single occurrence or in multiple occurrences, thus comprising one, two, three or more elements or characteristics, without any upper limit specified.

[0230] On the other hand, when an element is "designed" to perform a particular function, this means that the element is created specifically for the purpose of performing that particular function.

[0231] The expression "specifically designed" refers to an intentional design, targeted from the outset for a particular function or use. This expression implies that the element or system has been deliberately developed with special technical characteristics to fulfill this function, going beyond mere functional suitability or capability. For example, a system "specifically designed" for processing strain measurements incorporates, from its design stage, particular technical characteristics for this task, such as algorithms optimized for analyzing strain signals, filters adapted to the characteristic frequencies of the mechanical phenomena to be measured, or a hardware architecture sized for the sequential or real-time processing of strain data.This specific design is manifested in the deliberate technical choices and specific optimizations that distinguish the system from a generic solution simply adaptable to this use. Furthermore, the specific design involves a thorough consideration of the constraints and requirements specific to the intended application, leading to optimized and dedicated technical solutions that could not be obtained by simply adapting an existing system.

[0232] However, depending on the needs and resources available, consideration may be given to using an existing element, which will be modified or adapted to fulfill this particular function, without requiring substantial modifications to the invention.

[0233] As regards the expression "all or part," it indicates flexibility in the selection or use of the elements or data mentioned. This expression means that the action or characteristic described may apply to the entire set of elements or data in question, or only to a selected portion thereof. The use of "all or part" thus allows for a wide range of possibilities, from full to partial use, without specifying a precise lower or upper limit as to the quantity or proportion concerned.

[0234] It should be noted that the examples provided throughout this description are presented for illustrative purposes only and are not intended to be limiting. These examples are intended to facilitate understanding of the invention by a person skilled in the art, by providing concrete illustrations of possible implementation.

[0235] However, the invention is not limited to these specific examples. Those skilled in the art will understand that these examples can be generalized, adapted, or modified to suit specific needs, technological advances, or particular constraints, without departing from the spirit of the invention. Thus, whenever an example is given, it should be interpreted as encompassing not only the specific example mentioned, but also all equivalent technical variants and alternatives that perform the same function or achieve the same objective within the context of the invention.

[0236] The invention can be the subject of numerous variations and applications other than those described above. In particular, unless otherwise indicated, the various structural and functional features of each particular embodiment described above should not be considered as combined and / or closely and / or inextricably linked to one another, but, on the contrary, as mere juxtapositions. Furthermore, the structural and / or functional features of the various embodiments described above may be the subject, in whole or in part, of any different juxtaposition or any different combination.

Claims

1. Demands System (100) for processing deformation measurements of a silo support foot (11) of a silo (10) for storing or mixing at least one bulk material, the system (100) being specifically designed to collect deformation measurements and isolate the component related to the weight of the contents for use in determining the fill level of the silo (10), the system (100) comprising, - at least one deformation sensor (110) which is mechanically coupled to a single silo support foot (11) and which is specifically designed to measure the deformation of the silo support foot (11) in response to the introduction or removal of the bulk material, the deformation sensor (110) comprising at least one first electronic circuit which is specifically designed to generate, at least one first measurement signal (111) which includes deformation values ​​and timestamp data which may be integrated into the first measurement signal (111) or transmitted separately,- at least one transient environmental phenomenon detection sensor (120) that is mechanically coupled to the silo support foot (11) or that is located near the silo support foot (11), - at least one first wireless communication means (130) specifically designed to transmit the first measurement signal (111) with its associated timestamp data, - at least one remote server (140) comprising, - at least one second wireless communication means (141) specifically designed to receive the first measurement signal (111) with its associated timestamp data, - at least one first processor (142), and - at least one first storage memory (143), in which, - the transient environmental phenomenon detection sensor (120) is specifically designed to measure at least one of the following quantities: an ambient temperature quantity and / or a quantity representative of sunlight conditions,the sensor for detecting transient environmental phenomena, (120) comprising at least one second electronic circuit which is specifically designed to generate, — at least one second measurement signal (121) which includes local environmental measurement data corresponding to the measured quantities and characteristics of transient environmental phenomena affecting the silo support base (11) due to its exposure to local environmental conditions, and — timestamp data that can be integrated into the second measurement signal (121) or transmitted separately, - the first wireless communication means (130) is further specifically designed to transmit the second measurement signal (121) with its associated timestamp data, - the second wireless communication means (141) is further specifically designed to receive the second measurement signal (121) with its associated timestamp data, - the first processor (142) is specifically designed to — for the strain sensors (110), apply at least one machine learning model, previously trained to process the data from one or more strain sensors (110), to generate time series describing the variations in strain related to environmental phenomena according to the installation conditions on the base of the silo (11), — the machine learning model receiving as input at least one matrix comprising measurement signal data with their associated timestamp data, data including deformation values ​​and their associated timestamps, and environmental values ​​and their associated timestamps, — for each strain sensor (110), sequentially or in real time exploit the time series generated by its specific machine learning model to isolate the different strain components, based on the corresponding timestamps, - the first storage memory (143) is specifically designed to store, for each strain sensor (110), — its specific, pre-trained machine learning model, and — the time series describing the deformation variations generated by its specific machine learning model.

2. System (100) according to claim 1, wherein the first processor (142) is specifically designed to apply time series describing the deformation variations generated by the machine learning model according to the following steps: - categorize the local environmental measurement data extracted from the second measurement signal (121) into predetermined time periods corresponding to predetermined ranges of variation; - determine categorization thresholds based on: - the statistical distribution of the local environmental measurement data in the training dataset, and / or - transient environmental phenomena affecting the silo support foot (11); - for each predetermined time period,— determine at least one statistical value representative of the deformation components for the period as a function of local environmental measurement data extracted from the second measurement signals with their associated timestamp data, — store, in the first storage memory (143), the statistical value in association with the local environmental measurement data and their timestamps, — generate deformation components specific to the position of the deformation sensor (110) on the silo support foot (11) using the statistical value representative of the deformation components for the corresponding time period.

3. System (100) according to claim 2, wherein the predetermined ranges of variation include, - a range of low variations, below a first predetermined threshold, - a range of median variations, between the first threshold and a second predetermined threshold and - a range of high variations, above the second predetermined threshold, wherein, the first and second thresholds are determined as categorization thresholds.

4. A system (100) according to any one of claims 1 to 3, wherein the first processor (142) is specifically designed to apply sequentially or in real time the process of isolating the strain components to the strain measurements according to the following steps: - for each new strain measurement received at a given time via the first measurement signal (111) with its associated timestamp data, - determine a range of variation of the local environmental measurement data over a predetermined time period including the last measurement time, - select, from the time series generated by the specific machine learning model, a time series corresponding to the local environmental measurement data extracted from the second measurement signal (121) with its associated timestamp data for the range of variation, - extract from the time series a strain component corresponding to,— the time of the new strain measurement determined from the timestamp data associated with the first measurement signal (111), and — the local environmental measurement data extracted from the second measurement signal (121) with its associated timestamp data for the determined range of variation, — apply the strain component to the strain measurement to obtain an isolated measurement.

5. System (100) according to any one of claims 1 to 4, wherein, when used, the machine learning model is specifically designed to: - receive as input at least one matrix comprising: - previous strain values ​​with their associated timestamp data, and - at least one exogenous variable corresponding to local environmental measurement data with their associated timestamp data, the exogenous variable being used as an indicator of transient environmental phenomena, - generate a time series describing the expected deformation variations under the environmental conditions corresponding to the input matrix.

6. A system (100) according to any one of claims 1 to 5, wherein, when using the machine learning model, - the transient environmental phenomenon detection sensor (120) is specifically designed to generate at least two distinct exogenous variables corresponding respectively to: — first local environmental measurement data extracted from the second measurement signal (121) with its associated timestamp data, and — second local environmental measurement data extracted from the second measurement signal (121) with its associated timestamp data, characteristic of seasonal cycles; - the machine learning model is specifically designed to: — receive as input at least one matrix comprising the first and second data,and — preferably generate a single time series describing the deformation variations linked to environmental phenomena corresponding to the deformation values, environmental data, and their respective timestamps contained in the input matrix comprising the first and second data, or alternatively, generate several distinct time series based on these same data.

7. A system (200) for training at least one machine learning model for processing deformation measurements of a silo support foot (11) of a silo (10) for storing or mixing at least one bulk material, the machine learning model(s) being configured to process deformation measurements from one or more sensors installed on the silo support foot (11), the system (200) being specifically designed to generate deformation components adapted to the sensor installation conditions for use in determining the fill level of the silo (10), the system (200) comprising, - at least one second processor (210) specifically designed for, — identify time periods during which deformation variations are primarily due to transient environmental phenomena, including, — the analysis of deformation variations from measurements acquired by at least one deformation sensor (110) which is mechanically coupled to the silo support foot (11), the deformation sensor (110) being configured to generate at least one first measurement signal (111) comprising deformation values ​​and timestamp data which can be integrated into the first measurement signal (111) or transmitted separately, — the detection of regular variation patterns corresponding to measurements acquired by at least one transient environmental phenomenon detection sensor (120) which is mechanically coupled to the silo support foot (11) or which is located near the silo support foot (11), the sensor being configured to generate at least one second measurement signal (121) comprising local environmental measurement data and time-stamping data which can be integrated into the second measurement signal (121) or transmitted separately, — to create a training dataset comprising, — the extraction of deformation data and their associated timestamp data from measurements acquired by the deformation sensor (110), — centering the deformation data to eliminate discrepancies between time periods, — concatenation of data while preserving its temporal consistency, — train the machine learning model on the dataset to generate time series describing the deformation variations, including, — the preparation of at least one input matrix comprising data extracted from measurement signals with their associated timestamp data, data including strain values ​​and their associated timestamps, and environmental values ​​and their associated timestamps, — the use of the input matrix to train the model to generate deformation component predictions that take into account, — transient environmental phenomena affecting the silo support foot (11) due to its exposure to local environmental conditions according to its specific position, and — specific mechanical stresses exerted on the sensor due to its position on the structure of the silo support foot (11), - at least one second storage memory (220) specifically designed to store, — the training dataset, — the machine learning model, and — the time series generated by the machine learning model.

8. System (200) according to claim 7, wherein the second processor (210) is specifically designed to identify time periods according to at least one of the following approaches, - detection of periods of inactivity of the silo, - detection of regular and cyclic variation patterns characteristic of transient environmental phenomena, and - analysis of deformation variations correlated with local environmental measurement data, wherein the time periods include periods of at least (24) consecutive hours during which the silo exhibits regular variations in the measurements of the deformation sensor (110) corresponding to seasonal cycles.

9. System (200) according to any one of claims 7 to 8, wherein the second processor (210) is specifically designed to: - categorize local environmental measurement data extracted from the second measurement signal (121) into predetermined time periods corresponding to predetermined ranges of variation; - determine categorization thresholds based on: - the statistical distribution of local environmental measurement data in the training dataset, and - transient environmental phenomena affecting the silo support foot (11); - for each predetermined time period, - determine at least one statistically representative value of the deformation components for the period. — store, in the second storage memory (220), the statistical value representative of the deformation components for the corresponding time period in association with the local environmental measurement data and their timestamps.

10. System (200) according to claim 9, wherein the predetermined ranges of variation include, - a range of low variations, below a first predetermined threshold, - a range of median variations, between the first threshold and a second predetermined threshold and - a range of high variations, above the second predetermined threshold, wherein, the first and second thresholds are determined as categorization thresholds.

11. A system (200) according to any one of claims 7 to 10, wherein the second processor (210) is specifically designed to periodically update the machine learning model according to the following steps: - detection of a significant change in local environmental measurement data over a predetermined period corresponding to a seasonal variation, including: - calculation of a difference between the current local environmental measurement data and that used for training the machine learning model, taking into account the associated timestamps; - comparison of the difference to a predetermined threshold; - when the difference exceeds the predetermined threshold, - automatic creation of a new training dataset comprising the deformation values ​​and their timestamps, and the local environmental measurement data and their timestamps.— retraining the machine learning model with the new dataset, and — generation of new time series.

12. System (200) according to any one of claims 7 to 11, wherein the machine learning model is a time series prediction model specifically designed to, - receive as input at least one matrix comprising, — deformation values ​​with their associated timestamp data, and — at least one exogenous variable corresponding to local environmental measurement data with their associated timestamp data, the exogenous variable being used as an indicator of transient environmental phenomena, - generate a time series describing the expected deformation variations under the environmental conditions corresponding to the input matrix.

13. System (200) according to claim 12, comprising the generation of the time series according to a two-phase process comprising, — a first phase of creating and training a distinct model for each category of values ​​of the exogenous variable, and — a second phase of sequentially or real-time exploitation of the pre-trained models, where the system (200) identifies the category of the current exogenous variable and uses the corresponding model to generate a time series describing the shape of the deformation variations related to environmental phenomena under those specific conditions.

14. System (200) according to any one of claims 7 to 13, wherein, - the transient environmental phenomenon detection sensor (120) is specifically designed to generate at least two distinct exogenous variables corresponding respectively to, — first local environmental measurement data extracted from the second measurement signal (121) with its associated timestamp data, and — second local environmental measurement data extracted from the second measurement signal (121) with its associated timestamp data, characteristic of seasonal cycles, - the machine learning model is specifically designed to, — receive as input at least one matrix comprising the first and second data, and — generate distinct time series as a function of the input matrix comprising the first and second data.

Images