Structural health monitoring method

Abstract

A structural health monitoring method comprising: applying a plurality of sensor nodes (2) to a structure, each sensor node (2) comprising an accelerometer and a controller unit; providing a network gateway (3) operationally connected to the sensor nodes (2). The accelerometer (21) acquires an acceleration signal (20) of the structure. The controller unit applies a short-time Fourier transform (22) to the acceleration signal (20) to provide a transformed acceleration signal. The controller unit encodes the transformed acceleration signal by a convolutional encoder (24) to provide an encoded transformed acceleration signal, the convolutional encoder (24) including a previous training (51). The controller unit computes a Mahalanobis distance (25) for the encoded transformed acceleration signal according to a mean vector (µ) and a covariance matrix (S). The controller unit compares the Mahalanobis distance (25) with a predetermined threshold distance (26), and in case of exceeding the predetermined threshold distance (26) the controller unit sends acceleration data to the gateway (3), for further anomaly detection.

Description

[0001] Title: Structural health monitoring method

[0002] DESCRIPTION

[0003] Field of the invention

[0004] The present invention relates to a structural health monitoring method involving a plurality of sensor nodes, operationally connected to a network gateway, and associated with a structure.

[0005] In general, the present invention relates to electric digital data processing for monitoring of structures.

[0006] The present invention belongs to the technical field of infrastructure monitoring (roads, bridges, viaducts) and monitoring of general civil structures, preferably monitoring based on wireless sensors.

[0007] Background art

[0008] With development of infrastructure, the number of bridges, tunnels and the like increases. These key infrastructures are crucial for smooth and reliable operation of the road network.

[0009] However, infrastructure management capability can swiftly become a bottleneck for the supervision efficiency of the road network.

[0010] Known infrastructure risk prediction relies on manual analysis of monitoring data, which is complex, slow, and does not facilitate realtime risk assessment or timely warning of potential abnormal risks linked to the infrastructure’s health.

[0011] Acceleration is an important monitoring item for bridge structure health monitoring and is a data type mainly used for vibration-based structural damage identification and health state evaluation.

[0012] Acceleration response of a bridge is typically generated by excitation of external load action, such as vehicles, wind, etc., and is obtained by sensing through acceleration sensors arranged at different structural elements of a bridge. The bridge’s service performance continuously reduces due to the comprehensive influence of internal and external effects during the service period.

[0013] Most of the known data anomaly detection methods are based on longterm statistical analysis of historical data, including acceleration data, according to predetermined threshold values to find data anomalies through deterministic program screening and discrimination.

[0014] Document CN111737909A relates to a structural health monitoring data exception identification method based on a space-time diagram convolutional network; the identification method comprises: carrying out space-time correlation modeling on structure monitoring data by utilizing a space-time diagram convolutional network capable of learning an adjacent matrix, hierarchically applying information of adjacent nodes of each order to data regression, and designing a corresponding network structure and an objective function penalty term; and a monitoring system is used to establish initial measured data as a training set, a network is trained and an adjacent matrix is acquired, subsequent measured data is input into the network and then a model residual error and a diagnosis index are calculated, and the diagnosis index and a key adjacent edge are combined to determine whether data abnormity originates from a sensor fault or structural variation.

[0015] Document CN1153298 12A relates to a road infrastructure abnormity monitoring method based on artificial intelligence, and the method comprises the steps: adding a sensor in a road infrastructure, so as to monitor and collect related data, and carrying out the monitoring of the abnormity of the road infrastructure according to a sensor measurement point layout principle; the data is classified by combining multiple factors so as to judge whether monitoring is carried out or not, and the data is analyzed and processed by adopting a neural network deep learning method.

[0016] Document CN113866455A relates to deep learning-based bridge acceleration monitoring data anomaly detection method, system and device; the method comprises the following steps of obtaining real bridge acceleration time history data; processing the acquired acceleration time history data to obtain a 9-dimensional acceleration characteristic matrix; inputting the 9-dimensional acceleration characteristic matrix into a pre-trained LSTM neural network model; and determining the abnormal type of the bridge acceleration monitoring data according to the output of the LSTM neural network model.

[0017] Document CN115683504A relates to a bridge acceleration monitoring data exception identification method based on multi-label classification, and the method comprises the steps: determining an exception label for an acceleration data exception multi-label classification task according to the exception classification of acceleration data; constructing an acceleration data exception multi-label classification data set; constructing a deep learning model for data exception multi-label classification, and training and evaluating the model; unlabeled mass acceleration monitoring data are input into the trained multi-label classification model, so that the abnormal existence states of normal data and different types of data in each data segment can be identified.

[0018] However, known methods have limitations and there is a need for more advanced methods for identifying possible abnormal data, especially optimized for working on wireless accelerometer nodes.

[0019] There is a need for efficient and automated data anomaly recognition methods that can handle massive amounts of data and identify multiple types of data anomalies.

[0020] It is an objective of the present invention to solve drawbacks of the prior art.

[0021] In particular, it is an object of the present invention to provide a structural health monitoring method which is improved with respect to the prior art.

[0022] It is a further object of the present invention to provide a structural health monitoring method which is more effective in detecting anomalies of the structure.

[0023] It is a further object of the present invention to provide a structural health monitoring method which is less burdensome in its operational requirements.

[0024] It is a further object of the present invention to provide a structural health monitoring method which is more effective in energy consumption when the signals are collected by wireless sensors

[0025] It is a further object of the present invention to provide a structural health monitoring method which is more effective in prolonging the operational life of sensor nodes.

[0026] Summary of the invention

[0027] These and other objects of the present invention are achieved by a structural health monitoring method based on wireless accelerometer sensors and convolutional autoencoders.

[0028] A structural health monitoring method is set out in the appended claims, which are an integral part of the present description.

[0029] According to a preferred embodiment, the structural health monitoring method comprises applying a plurality of sensor nodes to a structure, each sensor node comprising at least one accelerometer and at least one controller unit; and providing a network gateway operationally connected to the plurality of sensor nodes.

[0030] The at least one accelerometer acquires an acceleration signal of the structure.

[0031] The at least one controller unit applies a short-time Fourier transform to the acceleration signal to provide a transformed acceleration signal.

[0032] The at least one controller unit encodes the transformed acceleration signal by a convolutional encoder to provide an encoded transformed acceleration signal, the convolutional encoder including a previous training. The at least one controller unit computes a Mahalanobis distance for the encoded transformed acceleration signal according to a mean vector and a covariance matrix.

[0033] The at least one controller unit compares the Mahalanobis distance with a predetermined threshold distance, and in case of exceeding the predetermined threshold distance the at least one controller unit sends acceleration data to the gateway, for further anomaly detection.

[0034] Advantageously, the structural health monitoring method allows detecting anomalies of the structure in an effective manner.

[0035] Advantageously, the structural health monitoring method is less burdensome in its operational requirements, since the majority of repetitive processing is performed directly by the sensor nodes.

[0036] Advantageously, the structural health monitoring method is more effective in energy consumption, since data is transmitted by the sensor nodes only in case of a possible anomaly.

[0037] Advantageously, the structural health monitoring method reduces the digital storage space required to store redundant acceleration signals, since anomalous accelerations only are sent to the gateway, or server, for further anomaly detection.

[0038] Other features and advantages of the invention will be apparent from the following description of preferred embodiments, and from the claims.

[0039] Brief description of the drawings

[0040] The invention will be now described with reference to the annexed drawings, provided as non-limiting examples of preferred embodiments, wherein:

[0041] Figure 1 exemplifies a system embodying the structural health monitoring method.

[0042] Figure 2 is a diagram exemplifying the training phase and operational phase of the structural health monitoring method.

[0043] Figure 3 is a further diagram exemplifying the training phase and operational phase of the structural health monitoring method.

[0044] In the drawings referred to in the description, the same reference numerals will designate the same or equivalent elements.

[0045] Detailed Description of the invention

[0046] Figure 1 exemplifies a system 1 embodying the structural health monitoring method, which will be further described.

[0047] The system 1 comprises a plurality of sensor nodes 2, which according to the structural health monitoring method are applied to a structure, such as a bridge.

[0048] The system 1 comprises a network gateway 3 which is operationally connected to the plurality of sensor nodes 2. With “operationally connected” it is intended that the sensor nodes 2 can at least transmit to, optionally can transmit to and receive from, the network gateway 3.

[0049] Preferably, the network gateway 3 is connected by a wireless connection to the plurality of sensor nodes 2, preferably by a Wi-Fi connection. Other suitable wireless connection protocol can be adopted.

[0050] The network gateway 3 is preferably operated by a Raspberry Pi, and serves as a pivotal component for data aggregation, management, and communication. The network gateway 3 acts as an MQTT broker, allowing the sensor nodes 2 to publish data on specific topics, facilitating communication between the sensor nodes and a central server. Furthermore, the network gateway 3 enables remote configuration of critical parameters such as sampling frequency and duty cycle of the sensor nodes 2.

[0051] Preferably, the gateway 3 connects to the Internet via an Ethernet cable available on the structure itself, and is powered directly by a cabled connection to the power grid. Preferably, the gateway 3 generates a WiFi network that enables wireless connectivity for the sensor nodes 2 acting as a hotspot generator. To make the connection accessible for every sensor node 2, one or more Wi-Fi repeaters can be added.

[0052] The system 1 preferably comprises at least one server 4 operationally connected to the network gateway 3. The at least one server 4 comprises at least one training unit, which will be further described. Once again, with “operationally connected” it is intended that the network gateway 3 can at least transmit to, preferably can transmit to and receive from, the at least one server 4.

[0053] Going back to the sensor nodes 2, each sensor node 2 comprises at least one accelerometer and at least one controller unit.

[0054] Preferably, the at least one accelerometer comprises a MEMS device. In particular, the MEMS accelerometer is selected to ensure a low-cost device. The main parameter that drives the selection of a MEMS accelerometer is its low noise density preferably having a value of 22.5 - , that is important when dealing with ambient (e.g. wind induced, traffic induced) vibration of civil structures. Furthermore, the selectable scale of a MEMS accelerometer ensures that an accurate measure within the range of accelerations can be performed, without saturating the signal or losing sensitivity. Finally, the selectable sampling frequency of a MEMS accelerometer allows adjustment based on the resonant frequencies of the concerned structure.

[0055] Preferably, the at least one controller unit comprises a programmable microcontroller. As it will be further described, the at least one controller unit operates in three distinct phases: data collection, data elaboration / processing and data transmission, with each process preferably occurring at separate times.

[0056] Preferably, each sensor node 2 further comprises at least a GPS module, so that synchronization between different sensor nodes 2 is provided by a GPS clock.

[0057] A GPS clock is used for synchronization between the different nodes. In particular, the starting time for acquisition is triggered at a given time by checking it from atomic clock from GPS. Then, in order to avoid clock drift problems, thanks to the selectable frequency of the time pulse provided by this GPS module, each acquisition point is triggered by means of an interrupt logic when the signal is high.

[0058] Preferably, each sensor node 2 further comprises at least a power management unit. In particular, each sensor node 2 is powered by a battery controlled by the at least a power management unit.

[0059] Preferably, each sensor node 2 further comprises at least one solar panel, which is preferably controlled by the at least a power management unit.

[0060] Preferably, the power management unit comprises a circuit which manages energy flow from the solar panel, to power the sensor while also regulating battery charging to maintain safe operating conditions. Preferably, such circuit employs a MPPT (Maximum Power Point Tracking) strategy to optimize energy harvesting from the solar panel, ensuring maximum power extraction. Simultaneously, such circuit controls the charging process to prevent overcharging or discharging of the battery, thus preserving its longevity, and ensuring stable operation of the sensor node 2. Finally, a step-down DC-DC converter is preferably used to lower the voltage of the battery to the value of 3.3V which is the exemplary operating voltage of the whole sensor node 2.

[0061] Figure 2 is a diagram exemplifying the training phase and operational phase of the structural health monitoring method.

[0062] In the operational phase 200, see right side of Figure 2, the at least one accelerometer 21 of sensor node 2 acquires an acceleration signal 20 of the structure.

[0063] With “an acceleration signal of the structure”, it is intended that the structure itself is subject to vibrations, which are induced by environmental (e.g., wind) or human (e.g. vehicular traffic) factors; the at least one accelerometer 21 acquires an acceleration signal 20 corresponding to the vibrations locally detected within the structure. As already explained, these vibrations of the structure are indicative of a health status of the structure itself.

[0064] The at least one controller unit of sensor node 2 applies a short-time Fourier transform 22 to the acceleration signal 20, to provide a transformed acceleration signal.

[0065] Preferably, the short-time Fourier transform 22 transforms the acceleration signal 20 from a one-dimensional vector in time-space domain into a two-dimensional array in frequency- time domain.

[0066] Preferably, the at least one controller unit of sensor node 2 applies the short-time Fourier transform with syncrosqueezing.

[0067] Preferably, the at least one controller unit of sensor node 2 is configured for scaling the transformed acceleration signal, according to a max absolute scaler rule 23 , as it will be further described.

[0068] The least one controller unit of sensor node 2 encodes the transformed acceleration signal by a convolutional encoder 24, to provide an encoded transformed acceleration signal. The convolutional encoder 24 includes a previous training, as it will be further described.

[0069] The at least one controller unit of sensor node 2 computes a Mahalanobis distance 25 for the encoded transformed acceleration signal, according to a mean vector ji and a covariance matrix S.

[0070] The at least one controller unit of sensor node 2 compares the computed Mahalanobis distance 25 with a predetermined threshold distance 26.

[0071] In case the computed Mahalanobis distance 25 exceeds the predetermined threshold distance 26, the at least one controller unit of sensor node 2 sends acceleration data to the gateway 3, for further anomaly detection

[0072] Preferably, the at least one server 4 provides the further anomaly detection of the acceleration data received from the gateway 3.

[0073] The acceleration data sent to the gateway 3 comprises the acceleration signal 20 and / or the transformed acceleration signal and / or the encoded transformed acceleration signal computed by the at least one controller unit of sensor node 2.

[0074] In case the computed Mahalanobis distance 25 does not exceed the predetermined threshold distance 26, the data 27 can be discarded, as the structure is within healthy operation.

[0075] In the training phase 100, see left side of Figure 2, the at least one accelerometer of sensor node 2 acquires an acceleration signal 10 of the structure.

[0076] In particular, the training phase 100 provides for acquiring the acceleration signal 10 of the structure in healthy conditions, preferably for at least thirty minutes daily for at least one month, more preferably for at least one hour daily and for at least six months.

[0077] Preferably, the at least one server 4 receives from the gateway 3 the acceleration signal 10 of the structure acquired by the at least one accelerometer of the sensor node 2 during the training phase 100.

[0078] The acceleration signal 10 thus corresponds to a training acceleration signal 10.

[0079] The at least one training unit of the at least one server 4 applies a short-time Fourier transform 42 to the training acceleration signal 10, to provide a transformed training acceleration signal.

[0080] Preferably, the at least one training unit of the at least one server 4 applies the short-time Fourier transform 42 with syncrosqueezing.

[0081] In case, the at least one training unit of the at least one server 4 is configured for scaling the transformed training acceleration signal, according to a max absolute scaler rule 43, in a manner compatible with the max absolute scaler rule 23 already described.

[0082] The at least one training unit of the at least one server 4 encodes the transformed training acceleration signal by the encoding portion 44 of a convolutional autoencoder 44 and 54, to provide an encoded transformed training acceleration signal.

[0083] The at least one training unit of the at least one server 4 further decodes the encoded transformed training acceleration signal by a decoding portion 54 of the same convolutional autoencoder 44 and 54, to provide a decoded transformed training acceleration signal.

[0084] As mentioned, the convolutional encoder 44 and the convolutional decoder 54 provide, together, a convolutional autoencoder 44 and 54.

[0085] An autoencoder is a type of artificial neural network used to learn efficient coding of unlabeled data, e.g. provide unsupervised learning.

[0086] In particular, the autoencoder 44, 54 learns two functions: an encoding function 44 that transforms the input data, and a decoding function 54 that recreates the input data from the encoded representation; the autoencoder 44, 54 learns an efficient representation (encoding) for the set of data.

[0087] In particular, the two convolutional networks which are comprised in the autoencoder 44 and 54 are trained simultaneously, so as to receive an input, compact the information, and reconstruct the input.

[0088] The at least one training unit of the at least one server 4 evaluates a loss function 53 between the encoded transformed training acceleration signal and the decoded transformed training acceleration signal.

[0089] The at least one training unit of the at least one server 4 minimizes the loss function 53 to provide training to a trained encoder 44 corresponding to a fraction of the convolutional autoencoder 44, 54.

[0090] The at least one training unit applies the trained encoder 44 to the transformed training acceleration signal to obtain the encoded transformed training acceleration signal, so as to determine the mean vector ji and the covariance matrix S, for example by fitting a multivariate Gaussian on the said encoded transformed training acceleration signal. The said trained encoder 44 is then set to be transferred to the convolutional encoder 24 of the at least one controller unit of sensor node 2.

[0091] More specifically, a training distribution 52 comprising the mean vector ji and the covariance matrix S is transferred, as item 51, to the at least one controller unit of the sensor node 2, used during the operational phase 200.

[0092] In particular, the nominal distribution 52 of data is represented by the mean vector |u and the covariance matrix S, which are transferred to the at least one controller unit of the sensor node 2.

[0093] Also, the trained encoder 44 is pushed to the at least one controller unit, so as to provide the encoder 24, which is used in the operational phase 200 for further analysis of the acceleration signal.

[0094] Figure 3 is a further diagram exemplifying the training phase and operational phase of the structural health monitoring method.

[0095] Figure 3 relates to the same embodiment of structural health monitoring method described with reference to Figure 2, with more details for further exemplification.

[0096] The algorithm applied for the structural health monitoring anomaly detection method based on acceleration measurements comprises a transformation process, a convolutional autoencoder, and an anomaly detection statistical module.

[0097] The acceleration signal 20 or 10 acquired by a single accelerometer node 2 is arranged in a vector sN, where N represents the number of samples.

[0098] A STFT-based synchrosqueezing transform 22 or 42 is applied to transform the initial time-space (lxN) into a frequency- time domain (FxT), where F is the number of frequencies and T the number of time steps considered for the synchrosqueezing transform. The transformation of a one-dimensional time vector to a two-dimensional array is deemed necessary because recurrent layers are not as effective as convolutional layers when it comes to compare training time, computational effort and capabilities in understanding complex relations.

[0099] The procedure is carried out in this way: the original acceleration signal 20 or 10 is reshaped into an array (SxNs) where S is the number of segments and Ns the samples per segment; for every segment the ST FT 22 or 42 is computed, with a time window of Nw samples and an overlapping of 50%. To guarantee the computation of square STFT, the following condition is imposed:

[0100] NW= 2^FS

[0101] All the STFTs are arranged in a 3-D array A of shape (SxFxF) being F=T. All the elements are scaled according to the max absolute scaler rule 23 or 43: where the subscripts i refers to the segment, j refers to the frequency, k to the time.

[0102] The S arrays are fed into a convolutional autoencoder 44 and 45, which is a particular type of neural network where the input and the output coincide, that is, the network is trained to reconstruct the input.

[0103] The hidden layer has a low-dimensional representation compared to the input and the output layer.

[0104] In an example, each layer is defined as follows (the dimension of every layer output tensor refers to the case where Nw =256, Ns= 16384, F=T= 129):

[0105] • Input layer: the input shape is equal to F x T x 1, (129, 129, 1).

[0106] • Convolutional 2-D layer: filters=32, Kernel size = 3x3. Activation=relu, (129, 129, 32).

[0107] • Max pooling 2-D layer: size=2x2, with padding (65, 65, 32). • Convolutional 2-D layer: filters=16, Kernel size = 3x3. Activation=relu, (65, 65, 16).

[0108] • Max pooling 2-D layer: size=2x2, with padding, (33, 33, 16).

[0109] • Convolutional 2-D layer: filters=8, Kernel size = 3x3. Activation=relu, (33, 33, 8).

[0110] • Max pooling 2-D layer: size=2x2, with padding, (17, 17, 8).

[0111] • Flatten layer: (2312).

[0112] • Dense layer: 8 neurons (8).

[0113] • Dense layer: same neurons as flatten layer (2312).

[0114] • Reshape layer: same shape as the last convolutional 2-D layer of the encoder (17, 17, 8).

[0115] • Convolutional 2-D transpose layer: filters=8, Kernel size = 3x3. Activation=relu, (17, 17, 8).

[0116] • Up-sampling 2-D layer: size=2x2, (34, 34, 8).

[0117] • Convolutional 2-D transpose layer: filters=16, Kernel size = 3x3. Activation=relu, (34, 34, 16).

[0118] • Up-sampling 2-D layer: size=2x2, (68, 68, 16).

[0119] • Convolutional 2-D transpose layer: filters=32, Kernel size = 3x3. Activation=relu, (68, 68, 32).

[0120] • Up-sampling 2-D layer: size=2x2, (136, 136, 32).

[0121] • Convolutional 2-D transpose layer: filters=l, Kernel size = 3x3. Activation=relu, (136, 136, 1).

[0122] • Cropping 2-D layer: cropping = (3, 4), (3, 4), (129, 129, 1).

[0123] In this example, the central hidden layer counts 8 neurons, then each 2-D array instance (FxF) is mapped into a 1x8 vector.

[0124] The loss function 53 utilized for the training of the autoencoder 44 and 45 is preferably the Mean Absolute Error (MAE), computed for a matrix of FxF elements as: being xtjthe prediction of the autoencoder 44, 54 in correspondence of the i-th row and the j -th column of the matrix.

[0125] The statistical anomaly detection module uses the Mahalanobis distance 25 to evaluate the distance between a new observation x (1x8) extracted from the hidden layer and a distribution, characterized by a vector of mean g (1x8) and a covariance matrix S (8x8). The Mahalanobis distance 25 is calculated as: dM= 7 (x - g)TS-1(x - g)

[0126] As already described, the structural health monitoring method passes through two separated phases: training phase 100 and operational phase 200.

[0127] In the training phase 100, a baseline period is defined, which accounts for accelerations acquired by the sensor nodes 2 from a structure considered in an “healthy” condition.

[0128] Preferably, the sensor nodes 2 acquire accelerations for one hour. Preferably, the sensing nodes 2 send the data to the gateway 3 via Wi-Fi once per hour. Preferably, the gateway 3 collects the data and send it to the server 4 one time per day.

[0129] All the computations associated with the training phase 100 are preferably carried out by the central server 4, in an off-line manner.

[0130] The STFT-based synchrosqueezing transform 42 is applied separately to data of every accelerometer sensor node 2. The 3-D arrays are scaled at item 43 separately for each accelerometer node sensor node 2. An autoencoder 44 and 54 is fitted separately for each accelerometer sensor node 2. For each accelerometer sensor node 2, all the vectors (1x8) extracted from the central hidden layer of the associated autoencoders 44 and 54 are used to define the distribution ( , S).

[0131] A statistical distribution 52 is therefore associated with every sensor node 2. Once all the steps are terminated, the trained model, the scaler and the distribution are pushed as item 51 to the corresponding sensor node 2.

[0132] The operational phase 200 works separately, as well, on each accelerometer sensor node 2, after the training of the autoencoders 44 and 54 and distributions 51 have been achieved on the server 4 and the models have been saved on each sensor node 2, in particular in the decoder 24 thereof.

[0133] The accelerations are acquired and saved in a sensor node 2 internal memory.

[0134] The calculations are carried out by the node controller unit.

[0135] The STFT-based synchrosqueezing transform 22 is applied to the acceleration vector (on a time frequency defined by the user).

[0136] The 3-D arrays extracted are scaled 23 with the max absolute scaler.

[0137] The resulting 3-D arrays are fed into the encoder 24, and the vectors (1x8) are extracted from the hidden layer.

[0138] The Mahalanobis distance 25 is computed for each vector.

[0139] If the Mahalanobis distance exceeds a threshold 26 defined by the user, the corresponding acceleration data sections are sent to the gateway 3. Otherwise, the acceleration data are deleted.

[0140] Industrial applicability

[0141] The structural health monitoring method of the present invention allows detecting anomalies of the structure in an effective manner and in a way which is optimized for wireless sensor nodes applied to a structure.

[0142] This invention allows to greatly reduce the energy consumption of the sensor nodes, being the data transmission the most energy-consuming process of the node. In this way, the non-critical acceleration signals are discarded, and only the relevant ones are transmitted to the gateway to be further analyzed, as well as the result of the algorithm applied on board.

[0143] Although the present invention has been described in considerable detail with reference to certain preferred embodiments thereof, other variants may become apparent to those skilled in the art who consider the present description.

[0144] For example, the duration and frequency of acceleration data acquisition, and data transmission can be further modified according to operational constraints and battery status of each node.

[0145] The embodiments shall be merely regarded as examples of the invention.

Claims

CLAIMS1. A structural health monitoring method comprising:- applying a plurality of sensor nodes (2) to a structure, each sensor node (2) comprising at least one accelerometer and at least one controller unit;- providing a network gateway (3) operationally connected to said plurality of sensor nodes (2); said at least one accelerometer (21) acquiring an acceleration signal (20) of said structure; said at least one controller unit applying a short-time Fourier transform (22) to said acceleration signal (20) to provide a transformed acceleration signal; said at least one controller unit encoding said transformed acceleration signal by a convolutional encoder (24) to provide an encoded transformed acceleration signal, said convolutional encoder (24) including a previous training (51); said at least one controller unit computing a Mahalanobis distance (25) for said encoded transformed acceleration signal according to a mean vector (p) and a covariance matrix (S); said at least one controller unit comparing said Mahalanobis distance (25) with a predetermined threshold distance (26), and in case of exceeding said predetermined threshold distance (26) said at least one controller unit sending acceleration data to said gateway (3), for further anomaly detection.

2. The structural health monitoring method according to claim 1, further comprising:- providing at least one server (4) operationally connected to said network gateway (3), said at least one server (4) comprising at least one training unit;said at least one server (4) receiving from said gateway (2) said acceleration signal of said structure acquired by said at least one accelerometer during a training phase (100), said acceleration signal corresponding to a training acceleration signal (10).

3. The structural health monitoring method according to claim 2, said at least one training unit applying a short-time Fourier transform (42) to said training acceleration signal (10) to provide a transformed training acceleration signal; said at least one training unit encoding said transformed training acceleration signal by a convolutional autoencoder (44, 54) to provide an encoded transformed training acceleration signal; said at least one training unit decoding said encoded transformed training acceleration signal by said convolutional autoencoder (44, 54), to provide a decoded transformed training acceleration signal; said at least one training unit evaluating a loss function (53) between said encoded transformed training acceleration signal and said decoded transformed training acceleration signal, and minimizing said loss function (53) to provide said previous training to a trained encoder (44) of said convolutional autoencoder (44, 54), said at least one training unit applying said trained encoder (44) to said transformed training acceleration signal to determine said mean vector (p) and said covariance matrix (S) said structural health monitoring method further comprising:- transferring (51) to said at least one controller unit a training distribution (52) comprising said mean vector (p) and said covariance matrix (S);- pushing (51) said trained encoder (44) to provide said convolutional encoder (24) in said at least one controller unit for further analysis of said acceleration signal (20).

4. The structural health monitoring method according to claim 3,said at least one training unit applying said short-time Fourier transform (42) with syncrosqueezing.

5. The structural health monitoring method according to claim 4, said at least one controller unit applying said short-time Fourier transform (22) with syncrosqueezing.

6. The structural health monitoring method according to any one of claims 2 to 5, said at least one server (4) providing said further anomaly detection of said acceleration data from said gateway (3).

7. The structural health monitoring method according to any one of claims 2 to 6, wherein said training phase (100) provides for acquiring said acceleration signal of said structure in healthy conditions, preferably for at least thirty minutes daily for at least one month.

8. The structural health monitoring method according to any one of claims 1 to 7, said structural health monitoring method further comprising:- scaling with said at least one controller unit said transformed acceleration signal, according to a max absolute scaler rule (23).

9. The structural health monitoring method according to any one of claims 1 to 8, wherein said short-time Fourier transform (22) transforms said acceleration signal (20) from a one-dimensional vector in time-space domain into a two-dimensional array in frequency-time domain.

10. The structural health monitoring method according to any one of claims 1 to 9, wherein said acceleration data comprises said acceleration signal (20) and / or said transformed acceleration signal and / or said encoded transformed acceleration signal.

11. The structural health monitoring method according to any one of claims 1 to 10, wherein said network gateway (3) is connected by a wireless connection to said plurality of sensor nodes (2), preferably by a Wi-Fi connection.

12. The structural health monitoring method according to any one of claims 1 to 11, wherein said each sensor node (2) further comprises at least a GPS module and wherein synchronization between different sensor nodes is provided by a GPS clock.

13. The structural health monitoring method according to any one of claims 1 to 12, wherein said each sensor node (2) further comprises at least a power management unit, wherein each sensor node is powered by a battery controlled by said at least a power management unit.

14. The structural health monitoring method according to claim 13, wherein said each sensor node (2) further comprises at least one solar panel controlled by said at least a power management unit.

15. The structural health monitoring method according to any one of claims 1 to 14, wherein said at least one accelerometer (21) comprises a MEMS device.

Images