A bearing multi-source sensing data acquisition method and system based on edge computing

Abstract

The application relates to the technical field of data processing, and specifically discloses a bearing multi-source sensing data acquisition method and system based on edge computing, which comprises the following steps: an edge computing node arranged in a bearing operation environment aligns a bearing environment parameter set and a bearing host running signal set received by the edge computing node according to timestamps, so that standardized time sequence data is obtained; a pre-constructed deep feature extraction network is used to take the standardized time sequence data as input, and a local correlation feature map between the environment parameters and the bearing host signal is extracted through a convolution layer; the local correlation feature map is input into a cycle layer, a dynamic evolution feature vector is extracted along a time axis, and an abnormality discrimination index is mapped through the edge computing node; and the health grade of the bearing operation environment of the bearing is determined according to the abnormality discrimination index; the application can reduce the transmission pressure of massive original data, reduce network bandwidth consumption and transmission time delay, and save cloud computing power resources.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a bearing multi-source sensor data acquisition method and system based on edge computing. Background Technology

[0002] Currently, industrial bearing operation status monitoring generally adopts a cloud-based centralized data acquisition and processing architecture. Multi-source sensing devices at the bearing operation site continuously generate massive amounts of raw monitoring data, all of which need to be uploaded to the cloud platform for analysis and judgment. Under this traditional architecture, the continuous uploading of massive amounts of raw data not only excessively consumes network bandwidth resources but also brings significant data transmission latency, failing to meet the actual needs of real-time bearing condition monitoring. It also results in the unnecessary consumption of cloud computing resources.

[0003] Existing bearing sensor data processing methods mostly employ independent data analysis models, severing the intrinsic connection between bearing environmental parameters and bearing host operating signals. This makes it difficult to uncover local correlation features among multi-source sensor data and dynamic evolution patterns over time. Conventional data acquisition and analysis methods lack on-site standardization and deep feature mining capabilities, resulting in insufficient accuracy in judging the health status of the bearing operating environment and an inability to achieve real-time anomaly assessment at the edge. Overall, the data acquisition efficiency and condition monitoring effect are insufficient to meet the high-precision, low-latency application requirements of industrial sites. Summary of the Invention

[0004] In view of the above problems, the purpose of this invention is to provide a bearing multi-source sensor data acquisition method and system based on edge computing, so as to solve the problems that existing bearing multi-source sensor data acquisition relies heavily on centralized cloud processing. Uploading all the massive amounts of raw sensor data from the field to the cloud will cause high network bandwidth consumption, high transmission latency, and waste cloud computing resources. In addition, there is a lack of timestamp alignment and standardization methods for bearing environmental parameters and bearing host operating signals, making it difficult to explore the local correlation features and temporal dynamic evolution features between multi-source data, and making it impossible to complete bearing anomaly identification and health level assessment at the field edge. The real-time performance and identification accuracy of bearing operating status monitoring cannot meet the needs of actual industrial applications.

[0005] This invention provides a bearing multi-source sensor data acquisition method based on edge computing, comprising:

[0006] Step 1: The edge computing nodes deployed in the bearing operating environment align the received bearing environment parameter set and bearing host operating signal set according to the timestamp to obtain the standardized time series data of the bearing.

[0007] Step 2: Based on the pre-constructed deep feature extraction network, using standardized time-series data within a continuous time window as input, extract the local correlation feature map between environmental parameters and bearing host signals through convolutional layers;

[0008] Step 3: Input the local correlation feature map into the recurrent layer, extract the dynamic evolution feature vector along the time axis, and map the node to the anomaly discrimination index through edge calculation;

[0009] Step 4: Determine the bearing operating environment health level based on the anomaly detection index, and upload the anomaly detection index and bearing environment health level to the remote management platform.

[0010] Preferably, the edge computing node deployed within the bearing operating environment aligns the received bearing environment parameter set and bearing host operating signal set according to timestamps to obtain standardized time-series data of the bearing. The process is as follows:

[0011] The edge computing node receives a set of bearing environmental parameters from a multi-source sensing acquisition device, as well as a set of bearing host operation signals from the bearing host monitoring module.

[0012] The edge computing node extracts the acquisition timestamp corresponding to each parameter value in the bearing environment parameter set, as well as the acquisition timestamp corresponding to each signal value in the bearing host operating signal set;

[0013] Using a fixed time interval as a reference, the bearing environmental parameter set and the bearing host operating signal set are resampled and aligned to obtain the bearing's multidimensional time series data matrix;

[0014] The multidimensional time series data matrix is ​​subjected to min-max normalization to map the numerical range of each dimension to a unified interval, thereby obtaining the standardized time series data of the bearing.

[0015] Preferably, the pre-built deep feature extraction network is constructed as follows:

[0016] The bearing environmental parameter set samples and bearing host operating signal set samples collected during historical periods are obtained, and after timestamp alignment and normalization, a historical training data matrix of the bearing is formed.

[0017] Training samples are extracted from the historical training data matrix using continuous time windows, and the corresponding bearing environmental health level is used as the training label.

[0018] Supervised learning algorithms are used to train the parameters of the convolutional layers in the deep feature extraction network, so as to minimize the prediction error between the local correlation feature maps output by the convolutional layers and the training labels.

[0019] Preferably, the process of using standardized time-series data within a continuous time window as input and extracting local correlation feature maps between environmental parameters and bearing host signals through convolutional layers is as follows:

[0020] Multiple parallel two-dimensional convolutional kernels are used, each with a preset size. The height of the kernel corresponds to the number of consecutive time steps within the time window, and the width corresponds to the total number of dimensions of the environmental parameters and the bearing host signal.

[0021] Each convolutional kernel is subjected to sliding convolution operation on the time dimension and parameter dimension of the standardized time series data to obtain the single-channel feature map corresponding to each convolutional kernel;

[0022] The single-channel feature maps corresponding to all convolutional kernels are concatenated along the channel dimension to form a local correlation feature map.

[0023] Preferably, the process of performing sliding convolution operations on the time and parameter dimensions of the standardized time-series data for each convolution kernel to obtain a single-channel feature map corresponding to each convolution kernel is as follows:

[0024] Each convolutional kernel slides along the time dimension according to a preset convolutional stride. After each slide, the inner product between the convolutional kernel and the standardized temporal data of the current coverage area is calculated to obtain the convolutional response value of the current coverage area.

[0025] The convolution response values ​​of each convolution kernel at all sliding positions are arranged in chronological order to form a single-channel feature map corresponding to each convolution kernel;

[0026] The single-channel feature map after nonlinear activation is used as the single-channel feature map corresponding to each convolution kernel.

[0027] Preferably, the process of inputting the local correlation feature map into the recurrent layer and extracting the dynamically evolving feature vector along the time axis is as follows:

[0028] The local correlation feature map is expanded according to a preset time step, and each time step corresponds to a local correlation feature vector.

[0029] Starting from the initial time step, the local correlation feature vectors of each time step are sequentially input into the recurrent layer of the deep feature extraction network;

[0030] The loop layer updates the hidden state of the current time step based on the input of the current time step and the hidden state of the previous time step;

[0031] The hidden state at the last time step is used as the dynamic evolution feature vector.

[0032] Preferably, the process of mapping edge computing nodes to anomaly detection indices is as follows:

[0033] The dynamically evolving feature vectors are input into the pre-configured fully connected layers in the edge computing nodes;

[0034] The fully connected layer performs a linear weighted summation of the dynamically evolving feature vectors to obtain the scalar values ​​of the dynamically evolving feature vectors;

[0035] The scalar value is input into the activation function, which maps the scalar value to a preset exponential range to obtain the bearing anomaly discrimination index.

[0036] Preferably, the mapping process of the anomaly detection index is as follows:

[0037] ;

[0038] In the formula, Scalar values ​​representing dynamically evolving eigenvectors. This represents the preset weight vector. This represents the transpose of the weight vector. This represents the preset bias scalar. This represents the dot product between the weight vector and the scalar values ​​of the dynamically evolving feature vector. Represents the natural constant.

[0039] Preferably, the process of determining the bearing operating environment health level based on the anomaly detection index and uploading the anomaly detection index and the bearing operating environment health level to the remote management platform is as follows:

[0040] The anomaly detection index is compared with a preset first threshold and a second threshold, wherein the first threshold is less than the second threshold;

[0041] If the anomaly discrimination index is less than or equal to the first threshold, the health level of the bearing operating environment is determined to be normal.

[0042] If the anomaly discrimination index is greater than the first threshold and less than the second threshold, the health level of the bearing operating environment is determined to be the warning level.

[0043] If the anomaly discrimination index is greater than or equal to the second threshold, the health level of the bearing operating environment is determined to be an alarm level.

[0044] Edge computing nodes encapsulate the anomaly detection index and the corresponding bearing operating environment health level into data packets, which are then sent to the receiving port of the remote management platform via wired or wireless networks.

[0045] The present invention also provides a bearing multi-source sensing data acquisition system based on edge computing, the system comprising:

[0046] The timing warping module is used by edge computing nodes deployed in the bearing operating environment to align the received bearing environment parameter set and bearing host operating signal set according to timestamps to obtain standardized timing data of the bearing.

[0047] The convolutional feature extraction module is used to extract local correlation feature maps between environmental parameters and bearing host signals by taking standardized time series data within a continuous time window as input, based on a pre-built deep feature extraction network;

[0048] The anomaly index mapping module is used to input the local correlation feature map into the recurrent layer, extract the dynamic evolution feature vector along the time axis, and map it into an anomaly discrimination index through edge calculation nodes;

[0049] The level reporting module is used to determine the bearing operating environment health level based on the anomaly discrimination index, and upload the anomaly discrimination index and bearing environment health level to the remote management platform.

[0050] As can be seen from the above technical solution, the present invention provides a bearing multi-source sensor data acquisition method and system based on edge computing. Relying on edge computing nodes, it completes the timestamp alignment and standardization of multi-source sensor data at the bearing operation site. Through a deep feature extraction network that integrates convolutional and recurrent layers, it mines the local correlation features and dynamic evolution features of the multi-source data over time. It can autonomously complete the calculation of anomaly discrimination index and the classification of bearing operating environment health levels at the edge, without needing to transmit massive amounts of raw sensor data to the cloud. This effectively reduces the pressure of massive data transmission, significantly reduces network bandwidth consumption and data transmission latency, and saves cloud computing resources. Simultaneously, it accurately captures the inherent correlation and temporal change patterns between bearing environmental parameters and bearing host operating signals, improving the accuracy and real-time performance of bearing health status judgment. Furthermore, it can upload the discrimination index and health level to a remote management platform, facilitating remote unified control of bearing operating status and better meeting the actual needs of real-time bearing monitoring and safe operation and maintenance in industrial sites. Attached Figure Description

[0051] Other objects and results of the invention will become more apparent and readily understood by referring to the following description taken in conjunction with the accompanying drawings, and with a more complete understanding of the invention. In the drawings:

[0052] Figure 1 This is a flowchart illustrating a bearing multi-source sensor data acquisition method based on edge computing according to an embodiment of the present invention.

[0053] Figure 2 This is a functional block diagram of a bearing multi-source sensor data acquisition system based on edge computing, according to an embodiment of the present invention. Detailed Implementation

[0054] Existing bearing multi-source sensor data acquisition largely relies on centralized cloud processing. Uploading all the massive amounts of raw sensor data from the field to the cloud results in high network bandwidth consumption, high transmission latency, and waste of cloud computing resources. Furthermore, there is a lack of methods for aligning and standardizing the timestamps of bearing environmental parameters and bearing host operating signals, making it difficult to uncover local correlation features and temporal dynamic evolution features between multi-source data. It is also impossible to complete bearing anomaly identification and health level assessment at the field edge. The real-time performance and accuracy of bearing operating status monitoring cannot meet the needs of actual industrial applications.

[0055] To address the aforementioned problems, this invention provides a bearing multi-source sensor data acquisition method and system based on edge computing. The specific embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0056] To illustrate the bearing multi-source sensing data acquisition method and system based on edge computing provided by this invention, Figure 1 An exemplary illustration is provided for a bearing multi-source sensor data acquisition method based on edge computing according to an embodiment of the present invention; Figure 2 An exemplary illustration is provided for a bearing multi-source sensor data acquisition system based on edge computing according to an embodiment of the present invention.

[0057] The following description of exemplary embodiments is merely illustrative and is in no way intended to limit the invention or its application or use. Techniques and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques and equipment should be considered part of the specification.

[0058] Reference Figure 1 The diagram shown is a flowchart illustrating a bearing multi-source sensor data acquisition method based on edge computing according to an embodiment of the present invention. In this embodiment, the bearing multi-source sensor data acquisition method based on edge computing includes:

[0059] Step 1: The edge computing nodes deployed in the bearing operating environment align the received bearing environment parameter set and bearing host operating signal set according to the timestamp to obtain the standardized time series data of the bearing.

[0060] In this embodiment of the invention, the edge computing node deployed within the bearing operating environment aligns the received bearing environment parameter set and bearing host operating signal set according to timestamps to obtain standardized time-series data of the bearing. The process is as follows:

[0061] The edge computing node receives a set of bearing environmental parameters from a multi-source sensing acquisition device, as well as a set of bearing host operation signals from the bearing host monitoring module.

[0062] The edge computing node extracts the acquisition timestamp corresponding to each parameter value in the bearing environment parameter set, as well as the acquisition timestamp corresponding to each signal value in the bearing host operating signal set;

[0063] Using a fixed time interval as a reference, the bearing environmental parameter set and the bearing host operating signal set are resampled and aligned to obtain the bearing's multidimensional time series data matrix;

[0064] The multidimensional time series data matrix is ​​subjected to min-max normalization to map the numerical range of each dimension to a unified interval, thereby obtaining the standardized time series data of the bearing.

[0065] Edge computing nodes receive data streams from multiple sensing devices via Ethernet or serial interfaces. Each data stream constitutes a bearing environmental parameter set, which is transmitted in fixed-format data frames. Each data frame contains a parameter type identifier, a parameter value in floating-point format, and a collection timestamp accurate to milliseconds. Simultaneously, the node receives data streams from the host monitoring module via another independent communication channel. These data streams constitute a host operation signal set, and their data frame format includes a signal type identifier, a signal value, and a millisecond-level timestamp. The node stores the received raw data frames into two independent buffers, forming the original bearing environmental parameter set and the host operation signal set.

[0066] Edge computing nodes parse the bearing environment parameter set in the buffer frame by frame, reading the timestamp field and parameter value field in each frame. The timestamp is converted from the original byte format into a unified time object that can be processed internally by the system, and this time object is bound to the parameter value of the frame and stored as a timestamp-parameter value pair. All pairs are arranged in ascending order of timestamp to form a time sequence. The same parsing operation is performed on the host operation signal set, extracting the timestamp and signal value in each frame, binding them into timestamp-signal value pairs and sorting them. Finally, two independent data sequences with time stamps are obtained, namely the bearing environment parameter sequence with timestamps and the host operation signal sequence with timestamps.

[0067] A fixed resampling time interval, such as one thousand milliseconds, is set. Starting from the earliest time point in two timestamped data sequences and ending at the latest time point, a series of equally spaced, millisecond-aligned target time points are generated. For each parameter dimension in the bearing environmental parameter sequence, at each target time point, the two closest actual data points in the parameter sequence are searched forward and backward. Based on the parameter values ​​and timestamps of these two actual data points, a weighted calculation is performed according to the time distance to deduce the parameter value at the target time point. The same search and weighted calculation process is performed on each signal dimension in the host operation signal sequence. Finally, all parameter dimensions and signal dimensions obtain the deduced values ​​at the same series of target time points. These values ​​are organized in chronological order, with each time point corresponding to a row of data and all dimensions as columns, thus generating a multidimensional time-series data matrix for the bearing.

[0068] Iterate through each column of the multidimensional time series data matrix, that is, each independent parameter or signal dimension, and sequentially compare all the values ​​in that column to find the largest and smallest values, which are recorded as the maximum and minimum values ​​of that column. For each original value in that column, perform the operation of subtracting the minimum value of that column, and then divide the result of the subtraction by the difference between the maximum and minimum values ​​of that column. In this way, each original value is transformed into a new value between zero and one. After all columns in the matrix have independently completed the above process of subtracting the minimum value and dividing by the difference, the value range of each dimension in the entire matrix has been uniformly adjusted to the interval between zero and one. The new matrix obtained at this time is the standardized time series data of the bearing.

[0069] Step 2: Based on the pre-constructed deep feature extraction network, using standardized time-series data within a continuous time window as input, extract the local correlation feature map between environmental parameters and bearing host signals through convolutional layers;

[0070] In this embodiment of the invention, the pre-built deep feature extraction network is constructed as follows:

[0071] The bearing environmental parameter set samples and bearing host operating signal set samples collected during historical periods are obtained, and after timestamp alignment and normalization, a historical training data matrix of the bearing is formed.

[0072] Training samples are extracted from the historical training data matrix using continuous time windows, and the corresponding bearing environmental health level is used as the training label.

[0073] Supervised learning algorithms are used to train the parameters of the convolutional layers in the deep feature extraction network, so as to minimize the prediction error between the local correlation feature maps output by the convolutional layers and the training labels.

[0074] The process of using standardized time-series data within a continuous time window as input and extracting local correlation feature maps between environmental parameters and bearing host signals through convolutional layers is as follows:

[0075] Multiple parallel two-dimensional convolutional kernels are used, each with a preset size. The height of the kernel corresponds to the number of consecutive time steps within the time window, and the width corresponds to the total number of dimensions of the environmental parameters and the bearing host signal.

[0076] Each convolutional kernel is subjected to sliding convolution operation on the time dimension and parameter dimension of the standardized time series data to obtain the single-channel feature map corresponding to each convolutional kernel;

[0077] The single-channel feature maps corresponding to all convolutional kernels are concatenated along the channel dimension to form a local correlation feature map.

[0078] The process of performing sliding convolution operations on the time and parameter dimensions of standardized time-series data for each convolution kernel to obtain the single-channel feature map corresponding to each convolution kernel is as follows:

[0079] Each convolutional kernel slides along the time dimension according to a preset convolutional stride. After each slide, the inner product between the convolutional kernel and the standardized temporal data of the current coverage area is calculated to obtain the convolutional response value of the current coverage area.

[0080] The convolution response values ​​of each convolution kernel at all sliding positions are arranged in chronological order to form a single-channel feature map corresponding to each convolution kernel;

[0081] The single-channel feature map after nonlinear activation is used as the single-channel feature map corresponding to each convolution kernel.

[0082] The system collects samples of all bearing environmental parameters acquired by sensors over a past period. These samples are time-series data sets containing multiple environmental parameters. Simultaneously, it collects samples of all bearing host operating signals acquired by the host monitoring module over the same historical period. These samples are time-series data sets containing multiple operating signals. For each collected sample, the system performs timestamp alignment processing according to the aforementioned steps, ensuring that the data from different sources within each sample strictly correspond at the time point. After alignment, each sample is then normalized to adjust the numerical range of all dimensions to the same interval. All processed samples are arranged sequentially according to their respective acquisition start times, with each sample as a data block. These data blocks are connected end-to-end in the time direction to form a large two-dimensional numerical table. This table is the bearing's historical training data matrix, with its rows corresponding to continuous time points from ancient times to the present, and its columns corresponding to different environmental parameters and host signal dimensions.

[0083] Define a continuous time window of fixed length, for example, containing one hundred consecutive sampling time points. Starting from the first row of the historical training data matrix, extract one hundred rows of data downwards. All columns in these one hundred consecutive rows constitute the first training sample. This training sample covers all parameter and signal changes from the first time point to the one hundredth time point. Simultaneously, consult the labeled bearing environmental health level records, which indicate the health status level of the bearing at each historical time point. Take the health level corresponding to the last time point covered by this training sample, i.e., the one hundredth time point, and use this health level value as the training label for the first training sample. Then, slide the time window downwards by a fixed step, for example, ten time points, and extract another one hundred rows of data starting from the eleventh row to form the second training sample. Take the health level of its last time point, i.e., the one hundred and tenth time point, as its training label. Repeat this sliding extraction and label matching process until the time window slides to the end of the historical training data matrix, thereby generating a dataset for supervised learning consisting of a large number of training samples and their corresponding training labels.

[0084] A gradient descent-based supervised learning algorithm is used to optimize each weight parameter of the convolutional layers in a deep feature extraction network. Specifically, a training sample generated in the previous step is input into the initial convolutional layer of the network. The convolutional layer calculates the value of its internal parameters based on the current value of the sample, outputting a result called a local correlation feature map. This feature map is fed into subsequent network layers, ultimately producing a predicted value for the bearing health level. This predicted value is compared with the sample's true training label, and a scalar error value is obtained using a specific error calculation formula. Then, the algorithm begins backpropagation, calculating the influence of each parameter on the total error, i.e., the gradient, layer by layer from the network output layer to the input layer. For each parameter in the convolutional layer, a gradient value is calculated. This gradient value indicates whether increasing the parameter will increase or decrease the error. Then, according to a pre-set small positive number, i.e., the learning rate, the gradient value of each parameter is multiplied by this learning rate. The current value of the parameter is then subtracted from this product to obtain the updated value of the parameter. This process is repeated multiple times for all training samples in the network. In each round, all training samples are used to update the parameters sequentially. Through this continuous and subtle parameter adjustment, the local correlation feature map output by the convolutional layer can guide the network to make predictions more and more accurately, ultimately minimizing the overall error between the predicted value and the true training label.

[0085] The convolutional layers of deep feature extraction networks contain multiple independent and structurally identical computational units called two-dimensional convolutional kernels. These convolutional kernels work in parallel, and each convolutional kernel has a predefined rectangular size. The height of its size explicitly determines how many consecutive time steps the convolutional kernel observes at once in the time dimension, and its width is strictly equal to the sum of the bearing environment parameters and the bearing host signal dimensions. This means that the width of each convolutional kernel is sufficient to cover all signal dimensions, so a single convolutional kernel is sufficient to capture the complete pattern within a specific time length across all data dimensions at once.

[0086] Each independent 2D convolutional kernel slides downwards along the time dimension with a preset, fixed-length convolution stride. It first covers the topmost rectangular data region of the training samples, which is the same size as itself. Then, it calculates each weight parameter stored inside it and multiplies it one-to-one with the corresponding value in the currently covered data region. All the multiplication results are summed to obtain a total. This total is the convolutional response value of the kernel at the current sliding position. After the calculation is completed, the convolutional kernel moves downwards according to the stride to cover the next rectangular data region of the same size as itself. The operation of multiplying and summing corresponding positions is repeated to generate the next convolutional response value. This process continues until the convolutional kernel covers the last possible data region at the bottom of the training samples, thus generating a convolutional response value for each possible sliding position.

[0087] For each independent convolutional kernel, all the convolutional response values ​​generated in the above sliding process in chronological order are arranged into a one-dimensional numerical sequence according to the order in which the sliding occurs, from the earliest time position to the latest time position. This one-dimensional numerical sequence, arranged in strict chronological order and generated by the convolutional kernel alone, constitutes the single-channel feature map corresponding to each convolutional kernel.

[0088] Each convolution response value in the single-channel feature map generated by each convolution kernel is input into a preset non-linear activation function for transformation. Taking the most commonly used modified linear unit function as an example, this function checks each input value. If the input value is greater than zero, it outputs the value directly. If the input value is less than or equal to zero, it outputs zero. Through this judgment and mapping process, each value in the original single-channel feature map is replaced by a new value that has undergone non-linear transformation. The new sequence formed after all values ​​have been replaced is the single-channel feature map corresponding to each convolution kernel after non-linear activation.

[0089] The single-channel feature maps generated by all parallel convolutional kernels after nonlinear activation are stitched together along the channel direction. Specifically, the single-channel feature map generated by the first convolutional kernel is used as the first channel plane of the output, the single-channel feature map generated by the second convolutional kernel is used as the second channel plane of the output, and so on. The feature map planes generated by all convolutional kernels are stacked together in order of their numbering to form a three-dimensional output data block with multiple channels. This three-dimensional data block containing feature information of multiple channels is the local correlation feature map calculated by the convolutional layer and ready to be output to the subsequent network layers.

[0090] Step 3: Input the local correlation feature map into the recurrent layer, extract the dynamic evolution feature vector along the time axis, and map the node to the anomaly discrimination index through edge calculation;

[0091] In this embodiment of the invention, the process of inputting the local correlation feature map into the recurrent layer and extracting the dynamically evolving feature vector along the time axis is as follows:

[0092] The local correlation feature map is expanded according to a preset time step, and each time step corresponds to a local correlation feature vector.

[0093] Starting from the initial time step, the local correlation feature vectors of each time step are sequentially input into the recurrent layer of the deep feature extraction network;

[0094] The loop layer updates the hidden state of the current time step based on the input of the current time step and the hidden state of the previous time step;

[0095] The hidden state at the last time step is used as the dynamic evolution feature vector.

[0096] The process of mapping edge computing nodes to anomaly detection indices is as follows:

[0097] The dynamically evolving feature vectors are input into the pre-configured fully connected layers in the edge computing nodes;

[0098] The fully connected layer performs a linear weighted summation of the dynamically evolving feature vectors to obtain the scalar values ​​of the dynamically evolving feature vectors;

[0099] The scalar value is input into the activation function, which maps the scalar value to a preset exponential range to obtain the bearing anomaly discrimination index.

[0100] The mapping process for the anomaly detection index is as follows:

[0101] ;

[0102] In the formula, Scalar values ​​representing dynamically evolving eigenvectors. This represents the preset weight vector. This represents the transpose of the weight vector. This represents the preset bias scalar. This represents the dot product between the weight vector and the scalar values ​​of the dynamically evolving feature vector. Represents the natural constant.

[0103] Local correlation feature maps are three-dimensional data structures output by convolutional layers. Their first dimension is the time step, representing the length of the feature map along the time axis. The second dimension is the number of feature channels, representing the number of different feature types extracted at each time step. The third dimension is usually a single value in two-dimensional convolution. Unfolding by a preset time step means slicing this three-dimensional structure along the time dimension. For each independent position in the time dimension, the values ​​of all feature channels corresponding to that position are extracted. These values ​​are arranged in ascending order of channel number and combined into a one-dimensional, continuous array of values. This one-dimensional array completely represents the local correlation features extracted by all convolutional kernels at that specific time step and is called the local correlation feature vector corresponding to that time step. After processing all time steps in sequence, a sequence of local correlation feature vectors arranged in chronological order is obtained.

[0104] Starting from the first position in the sequence, i.e., the initial time step, the local correlation feature vector corresponding to that time step is extracted. This vector is an array containing multiple floating-point numbers. It is used as the input data for the current time step and fed completely into the input of the pre-set recurrent layer unit in the deep feature extraction network. Then, the sequence moves to the next time step, extracts the local correlation feature vector of the next time step, and feeds it completely into the same recurrent layer unit. This process is carried out strictly in chronological order from early to late. The recurrent layer unit processes each input vector in turn until the local correlation feature vector of the last time step in the sequence is also input and processed, thus ensuring that the integrity and order of the time series information are perceived by the recurrent layer unit.

[0105] When the recurrent layer receives the local correlation feature vector input at the current time step, it combines it with an internal state called the hidden state vector for computation. The hidden state vector is a one-dimensional array with a different dimension than the input vector, encoding all historical information up to the previous time step. For the initial time step, since there is no previous time step, the hidden state vector of the previous time step is initialized to a vector with all elements equal to zero. The core of the recurrent layer's computation is to concatenate the input vector of the current time step and the hidden state vector of the previous time step along the dimensional direction, forming a longer composite vector. Then, a linear transformation operation is performed on this composite vector, which involves multiplying each element of the composite vector with a set of preset weight parameters and summing the results, plus a set of preset bias parameters. Finally, the result of the linear transformation is input into a non-linear activation function, which maps each input value to a new range. The result of this computation process is the updated hidden state vector at the current time step, which contains the sequence information up to the current moment and is immediately stored as the hidden state vector of the previous time step when processing the input of the next time step.

[0106] After the recurrent layer unit processes the local correlation feature vector of the last time step in the input sequence in sequence, it performs the same internal calculation as the previous time step and outputs the updated hidden state vector of the final time step. This vector is the final internal state obtained by the recurrent layer after recursive calculation from the first to the last time step. It condenses and encodes all the dynamic evolution information of the entire input sequence from the initial time step to the last time step. The hidden state vector of the last time step is completely extracted from the output of the recurrent layer unit without any further time recursion. This vector is defined as the summary output of the entire recurrent layer processing result and is called the dynamic evolution feature vector.

[0107] The dynamic evolution feature vector is a one-dimensional floating-point array. As a high-level abstract feature extracted from time-series data, it is directly transmitted to a pre-configured fully connected layer processing module deployed on edge computing nodes. Edge computing nodes refer to embedded or dedicated computing devices close to the data acquisition source. The fully connected layer is a fixed structure part of the program model running on the device, and its internal parameters have been determined and solidified during the training phase.

[0108] The fully connected layer is pre-configured to contain only one neuron, which has a set of weights with the exact same dimensions as the dynamically evolving feature vector. Each weight is an adjustable floating-point number. The neuron also has a separate bias value. During computation, the first element of the dynamically evolving feature vector is multiplied by the first weight of the neuron, the second element is multiplied by the second weight, and so on, until all elements are multiplied by their corresponding weights. All these products are then added together to obtain a cumulative sum. Finally, this cumulative sum is added to the neuron's bias value to obtain a single, comprehensive numerical sum. This numerical sum is the scalar value of the dynamically evolving feature vector after the linear transformation by the fully connected layer.

[0109] The scalar value is input into a preset, specific activation function for final mapping. The sigmoid function is used, which calculates by first taking the negative of the scalar value and using this negative number as the exponent of the natural constant e to obtain the result of e raised to the power of the negative scalar value. Then, the result is added to the number one. Finally, the reciprocal of the new result is taken, which is the result of dividing by one. Through this series of defined exponentiation and division operations, any input scalar value will be mapped and compressed into a continuous numerical range between zero and one. The final output value between zero and one is the bearing anomaly discrimination index, which is used to directly indicate the degree of abnormality in the bearing condition.

[0110] The scalar value of the dynamically evolving feature vector is obtained by performing a global aggregation operation on the multidimensional hidden states output by the recurrent layers in the deep feature extraction network. Specifically, after the recurrent layer completes the recursive calculation of the locally correlated feature vectors for all time steps, it outputs a multidimensional vector containing rich temporal dynamic information. At this point, a global averaging aggregation algorithm is used to process this multidimensional vector, which involves summing the values ​​at all dimensions of the vector and then dividing the sum by the total number of dimensions of the vector itself to calculate a single average value. This value is the scalar value of the dynamically evolving feature vector, which highly summarizes the complete operational state evolution trajectory of the bearing from its initial state to the current moment.

[0111] The weight vectors are obtained by training a deep neural network using a large amount of historical monitoring data throughout the bearing's entire lifecycle. During the offline training phase, the preprocessed historical training data matrix is ​​input into the deep feature extraction network, and the network parameters are continuously optimized using a supervised learning algorithm. The specific optimization process involves inputting training samples into the network, calculating the error between the network's predicted output and the true label, then using the backpropagation algorithm to calculate the gradient, and continuously fine-tuning the weight vector values ​​in the opposite direction of the gradient. After multiple rounds of iterative training, the network can accurately identify the bearing's health status. The converged set of weight values ​​is then fixed and pre-set into the fully connected layer of the edge computing nodes for weighted evaluation of the dynamically evolving feature vectors during the online phase.

[0112] The transpose of the weight vector is a standard mathematical dimensionality adjustment operation performed during linear transformation calculations in fully connected layers to accommodate the scalar dimension of the dynamically evolving feature vector. Since the dynamically evolving feature vector is a single-dimensional scalar value, while the weight vector is typically defined as a column vector with the same dimension, the row and column orientations of the weight vector must be interchanged to enable subsequent multiplication and addition operations. Specifically, the vertical arrangement of the weight vector is adjusted to a horizontal arrangement, generating a transposed weight vector in row vector form, thus ensuring consistency and validity in mathematical calculation with the scalar value.

[0113] The bias scalar is also obtained based on the historical bearing lifecycle monitoring dataset, trained using the same supervised learning algorithm and backpropagation optimization mechanism mentioned above. During the training of the deep neural network, the bias scalar, as an adjustment parameter independent of the input features in the fully connected layer, participates in error calculation and gradient backpropagation synchronously with the weight vector. Its role is to provide a basic baseline offset for the model's linear output, eliminating the systematic bias caused by the model relying solely on the multiplication of weights and inputs. After sufficient offline iterative training, the bias value that enables the anomaly detection model to achieve optimal prediction results is ultimately determined as the preset bias scalar and deployed in the fully connected layer of the edge computing nodes for fixed use.

[0114] The natural constant is a fundamental physical constant in mathematics with a fixed irrational number value, approximately 2.71828. This constant is widely defined and applied in numerous fields of mathematical analysis and natural sciences, representing the fundamental rate limit of continuous compound growth or decay in nature. In the mapping process of this discriminant exponent, the natural constant, as an immutable objective mathematical benchmark, is directly used as the core base input to a specific nonlinear activation function for exponential operations. Its specific value is pre-set and constant within the relevant mathematical computational environment.

[0115] The core significance of this mapping process lies in transforming the high-dimensional abstract features extracted from the deep neural network and linearly weighted by fully connected layers into a standardized anomaly detection index with clear physical meaning, ranging from zero to one. Specifically, the dynamically evolving feature vector, after undergoing a transpose operation of the weight vector and an addition operation with a preset bias scalar, forms a linear combination value that comprehensively reflects the current multidimensional state characteristics of the bearing. Subsequently, this linear combination value is substituted into the logistic function for nonlinear compression mapping. Through the mathematical properties of exponential operations, the linear value, which might originally span an infinite range of positive and negative values, is smoothly constrained and distributed within a finite interval of zero to one. This final continuous value between zero and one is the anomaly detection index, which can intuitively and quantitatively indicate the degree of anomaly in the bearing's current deviation from normal operating conditions. The closer the value is to one, the higher the degree of anomaly; the closer it is to zero, the healthier the condition.

[0116] As the sum of the weight vector transpose, the scalar value of the dynamically evolving feature vector, and the preset bias scalar increases towards positive infinity, the anomaly discrimination index calculated through this mapping process exhibits a monotonically increasing trend, gradually approaching the value of one. Conversely, as the sum of the above three values ​​decreases towards negative infinity, the anomaly discrimination index exhibits a monotonically decreasing trend, gradually approaching the value of zero. Throughout the entire real number range, this mapping process maintains a smooth and continuous monotonically increasing characteristic, ensuring that any minute changes in the bearing state are continuously and non-linearly amplified and reflected in the anomaly discrimination index, thereby achieving highly sensitive capture and accurate quantitative characterization of early, subtle anomalies in the bearing.

[0117] Step 4: Determine the bearing operating environment health level based on the anomaly detection index, and upload the anomaly detection index and bearing environment health level to the remote management platform.

[0118] In this embodiment of the invention, the process of determining the bearing operating environment health level based on the anomaly detection index and uploading the anomaly detection index and the bearing operating environment health level to the remote management platform is as follows:

[0119] The anomaly detection index is compared with a preset first threshold and a second threshold, wherein the first threshold is less than the second threshold;

[0120] If the anomaly discrimination index is less than or equal to the first threshold, the health level of the bearing operating environment is determined to be normal.

[0121] If the anomaly discrimination index is greater than the first threshold and less than the second threshold, the health level of the bearing operating environment is determined to be the warning level.

[0122] If the anomaly discrimination index is greater than or equal to the second threshold, the health level of the bearing operating environment is determined to be an alarm level.

[0123] Edge computing nodes encapsulate the anomaly detection index and the corresponding bearing operating environment health level into data packets, which are then sent to the receiving port of the remote management platform via wired or wireless networks.

[0124] The central processing unit (CPU) of the edge computing node executes a read instruction to read two pre-programmed fixed values ​​from a specific storage sector of a non-volatile memory chip, such as flash memory, on the node's motherboard. These two values ​​are artificially set based on statistical analysis of a large amount of historical bearing operating data and are named the first threshold and the second threshold, respectively. The specific value of the first threshold is set to be less than the specific value of the second threshold. After reading, the CPU retrieves the specific value of the anomaly discrimination index, which represents the current bearing state evaluation result and was just calculated by the fully connected layer and activation function, from its internal register dedicated to storing calculation results. Subsequently, the CPU calls the numerical comparator circuit in its arithmetic logic unit. This circuit performs two comparison operations simultaneously by performing two subtraction operations in parallel and checking the sign bit of the result. The first comparison subtracts the value of the anomaly discrimination index from the value of the first threshold, and the second comparison subtracts the value of the anomaly discrimination index from the value of the second threshold. Based on the positive or negative sign and whether the result is zero generated by the two subtraction operations, the system can accurately and synchronously determine the relationship between the anomaly discrimination index and the two preset thresholds, which is less than, equal to, or greater than the two thresholds, thus completing the entire comparison process.

[0125] When the status flag output by the arithmetic logic unit indicates that the result of the anomaly discrimination index minus the first threshold is less than or equal to zero, the logic judgment unit responsible for program flow control in the edge computing node, usually the instruction control unit embedded in the central processing unit, will jump to the corresponding code segment according to the preset conditional branch instruction. The operation performed by this code segment is to copy a predefined string constant representing the "normal" state, such as "Normal_Level", or a corresponding integer status code, such as the number zero, from the program's read-only data area to a variable address space in the system memory designated to store the health level result. This written string or status code is defined as the normal level of the bearing's operating environment health level. It is stored as a clear qualitative judgment conclusion to indicate that the bearing is currently in a completely healthy operating state and requires no maintenance or intervention.

[0126] When the status flag output by the arithmetic logic unit indicates that the result of the anomaly discrimination index minus the first threshold is greater than zero, and at the same time the result of the anomaly discrimination index minus the second threshold is less than zero, the logic decision unit will jump to another different code segment according to the conditional branch instruction. The operation performed by this code segment is to copy a predefined string constant representing the "warning" status, such as "Warning_Level", or a corresponding integer status code, such as the number one, from the read-only data area to the memory variable address storing the health level result. This written string or status code is defined as the warning level of the bearing operating environment health level. It is stored as a clear qualitative judgment conclusion to indicate that the bearing's operating status has deviated from the baseline of complete health, and there are slight abnormal signs. It requires attention from equipment operators or maintenance systems and may arrange planned inspections, but it has not yet reached the severity level that requires immediate shutdown.

[0127] When the status flag output by the arithmetic logic unit indicates that the result of the anomaly discrimination index minus the second threshold is greater than or equal to zero, the logic decision unit will jump to the third code segment according to the conditional branch instruction. The operation performed by this code segment is to copy a predefined string constant representing the "alarm" status, such as "Alarm_Level", or a corresponding integer status code, such as the number two, from the read-only data area to the memory variable address storing the health level result. This written string or status code is defined as the alarm level of the bearing operating environment health level. It is stored as a clear qualitative judgment conclusion to indicate that the bearing's operating status has experienced a serious anomaly and there is a high risk of failure. The system must immediately trigger the shutdown protection mechanism or issue the highest priority maintenance notice to prevent equipment damage or safety accidents.

[0128] The application running on the edge computing node calls its communication protocol stack functions. First, it reads two key data items from system memory: a floating-point value representing the anomaly detection index (a quantitative assessment value) and a string or status code representing the bearing operating environment health level (a qualitative judgment). Then, following a pre-defined binary structure (e.g., using JavaScript object representation or protocol buffer format), it assembles this data into an ordered data body. At the beginning of the data body, the program adds a header structure, which sequentially fills in the current system's precise timestamp, the edge computing node's unique device identifier, and a representation of the "bearing health status." The data type encoding of the "report" is at the end of the data body. The program calculates the cyclic redundancy check (CRC) code for the entire message header and data body and appends the CRC code to the end of the message, thus forming a complete data message from head to tail. Subsequently, the communication protocol stack calls the transmission control protocol or user datagram protocol's send function to encapsulate the data message as a payload into a data packet at the network layer. Finally, through the wired Ethernet controller or wireless cellular network module integrated on the node hardware, the data packet is converted into an electrical signal or electromagnetic wave signal and sent through the physical network link to the network receiving port identified by the specific Internet Protocol address and transport layer port number published by the remote management platform server.

[0129] As can be seen from the above embodiments, the bearing multi-source sensor data acquisition method based on edge computing provided by the present invention completes the timestamp alignment and standardization of multi-source sensor data at the bearing operation site by relying on edge computing nodes. By fusing convolutional and recurrent layers through a deep feature extraction network, it mines the local correlation features and dynamic evolution features of the time dimension of multi-source data. It can autonomously complete the calculation of anomaly discrimination index and the classification of bearing operation environment health level at the edge, without the need to transmit massive amounts of raw sensor data to the cloud, effectively reducing the pressure of massive data transmission, significantly reducing network bandwidth consumption and data transmission latency, saving cloud computing resources. At the same time, it accurately captures the inherent correlation and temporal change law between bearing environmental parameters and bearing host operating signals, improving the accuracy and real-time performance of bearing health status judgment. It can also upload the discrimination index and health level to a remote management platform, which facilitates the remote unified control of bearing operation status and better meets the actual needs of real-time monitoring and safe operation and maintenance of bearings in industrial sites.

[0130] like Figure 2 The diagram shown is a functional block diagram of a bearing multi-source sensor data acquisition system 100 based on edge computing provided in an embodiment of the present invention, including a time-series warping module 101, a convolutional feature extraction module 102, an anomaly index mapping module 103, and a level reporting module 104.

[0131] In this embodiment, the functions of each module are as follows:

[0132] The timing warping module 101 is used by the edge computing nodes deployed in the bearing operating environment to align the received bearing environment parameter set and bearing host operating signal set according to the timestamp to obtain the standardized timing data of the bearing.

[0133] The convolutional feature extraction module 102 is used to extract local correlation feature maps between environmental parameters and bearing host signals through convolutional layers, based on a pre-constructed deep feature extraction network and standardized time-series data within a continuous time window as input.

[0134] The anomaly index mapping module 103 is used to input the local correlation feature map into the recurrent layer, extract the dynamic evolution feature vector along the time axis, and map it into an anomaly discrimination index through edge computing nodes;

[0135] The level reporting module 104 is used to determine the bearing operating environment health level of the bearing based on the anomaly discrimination index, and upload the anomaly discrimination index and the bearing environment health level to the remote management platform.

[0136] As can be seen from the above embodiments, the bearing multi-source sensor data acquisition system based on edge computing provided by the present invention completes the timestamp alignment and standardization of multi-source sensor data at the bearing operation site by relying on edge computing nodes. By fusing convolutional and recurrent layers through a deep feature extraction network, it mines the local correlation features and dynamic evolution features of the time dimension of multi-source data. It can autonomously complete the calculation of anomaly discrimination index and the classification of bearing operation environment health level at the edge, without transmitting massive amounts of raw sensor data to the cloud, effectively reducing the pressure of massive data transmission, significantly reducing network bandwidth consumption and data transmission latency, saving cloud computing resources. At the same time, it accurately captures the inherent correlation and temporal change law between bearing environmental parameters and bearing host operating signals, improving the accuracy and real-time performance of bearing health status judgment. It can also upload the discrimination index and health level to a remote management platform, which facilitates the remote unified control of bearing operation status and better meets the actual needs of real-time monitoring and safe operation and maintenance of bearings in industrial sites.

[0137] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A bearing multi-source sensor data acquisition method based on edge computing, characterized in that, The method includes: Step 1: The edge computing nodes deployed in the bearing operating environment align the received bearing environment parameter set and bearing host operating signal set according to the timestamp to obtain the standardized time series data of the bearing. Step 2: Based on the pre-constructed deep feature extraction network, using standardized time-series data within a continuous time window as input, extract the local correlation feature map between environmental parameters and bearing host signals through convolutional layers; Step 3: Input the local correlation feature map into the recurrent layer, extract the dynamic evolution feature vector along the time axis, and map the node to the anomaly discrimination index through edge calculation; Step 4: Determine the bearing operating environment health level based on the anomaly detection index, and upload the anomaly detection index and bearing environment health level to the remote management platform.

2. The bearing multi-source sensor data acquisition method based on edge computing as described in claim 1, characterized in that, The edge computing nodes deployed within the bearing operating environment align the received bearing environment parameter set and bearing host operating signal set according to timestamps to obtain standardized time-series data of the bearing. The process is as follows: The edge computing node receives a set of bearing environmental parameters from a multi-source sensing acquisition device, as well as a set of bearing host operation signals from the bearing host monitoring module. The edge computing node extracts the acquisition timestamp corresponding to each parameter value in the bearing environment parameter set, as well as the acquisition timestamp corresponding to each signal value in the bearing host operating signal set; Using a fixed time interval as a reference, the bearing environmental parameter set and the bearing host operating signal set are resampled and aligned to obtain the bearing's multidimensional time series data matrix; The multidimensional time series data matrix is ​​subjected to min-max normalization to map the numerical range of each dimension to a unified interval, thereby obtaining the standardized time series data of the bearing.

3. The bearing multi-source sensor data acquisition method based on edge computing as described in claim 1, characterized in that, The pre-built deep feature extraction network is constructed as follows: The bearing environmental parameter set samples and bearing host operating signal set samples collected during historical periods are obtained, and after timestamp alignment and normalization, a historical training data matrix of the bearing is formed. Training samples are extracted from the historical training data matrix using continuous time windows, and the corresponding bearing environmental health level is used as the training label. Supervised learning algorithms are used to train the parameters of the convolutional layers in the deep feature extraction network, so as to minimize the prediction error between the local correlation feature maps output by the convolutional layers and the training labels.

4. The bearing multi-source sensor data acquisition method based on edge computing as described in claim 3, characterized in that, The process of using standardized time-series data within a continuous time window as input and extracting local correlation feature maps between environmental parameters and bearing host signals through convolutional layers is as follows: Multiple parallel two-dimensional convolutional kernels are used, each with a preset size. The height of the kernel corresponds to the number of consecutive time steps within the time window, and the width corresponds to the total number of dimensions of the environmental parameters and the bearing host signal. Each convolutional kernel is subjected to sliding convolution operation on the time dimension and parameter dimension of the standardized time series data to obtain the single-channel feature map corresponding to each convolutional kernel; The single-channel feature maps corresponding to all convolutional kernels are concatenated along the channel dimension to form a local correlation feature map.

5. The bearing multi-source sensor data acquisition method based on edge computing as described in claim 4, characterized in that, The process of performing sliding convolution operations on the time and parameter dimensions of standardized time-series data for each convolution kernel to obtain the single-channel feature map corresponding to each convolution kernel is as follows: Each convolutional kernel slides along the time dimension according to a preset convolutional stride. After each slide, the inner product between the convolutional kernel and the standardized temporal data of the current coverage area is calculated to obtain the convolutional response value of the current coverage area. The convolution response values ​​of each convolution kernel at all sliding positions are arranged in chronological order to form a single-channel feature map corresponding to each convolution kernel; The single-channel feature map after nonlinear activation is used as the single-channel feature map corresponding to each convolution kernel.

6. The bearing multi-source sensor data acquisition method based on edge computing as described in claim 5, characterized in that, The process of inputting the local correlation feature map into the recurrent layer and extracting the dynamically evolving feature vector along the time axis is as follows: The local correlation feature map is expanded according to a preset time step, and each time step corresponds to a local correlation feature vector. Starting from the initial time step, the local correlation feature vectors of each time step are sequentially input into the recurrent layer of the deep feature extraction network; The loop layer updates the hidden state of the current time step based on the input of the current time step and the hidden state of the previous time step; The hidden state at the last time step is used as the dynamic evolution feature vector.

7. The bearing multi-source sensor data acquisition method based on edge computing as described in claim 6, characterized in that, The process of mapping edge computing nodes to anomaly detection indices is as follows: The dynamically evolving feature vectors are input into the pre-configured fully connected layers in the edge computing nodes; The fully connected layer performs a linear weighted summation of the dynamically evolving feature vectors to obtain the scalar values ​​of the dynamically evolving feature vectors; The scalar value is input into the activation function, which maps the scalar value to a preset exponential range to obtain the bearing anomaly discrimination index.

8. The bearing multi-source sensor data acquisition method based on edge computing as described in claim 7, characterized in that, The mapping process for the anomaly detection index is as follows: ； In the formula, Scalar values ​​representing dynamically evolving eigenvectors. This represents the preset weight vector. This represents the transpose of the weight vector. This represents the preset bias scalar. This represents the dot product between the weight vector and the scalar values ​​of the dynamically evolving feature vector. Represents the natural constant.

9. The bearing multi-source sensor data acquisition method based on edge computing as described in claim 1, characterized in that, The process of determining the bearing's operating environment health level based on the anomaly detection index and uploading the anomaly detection index and the bearing's operating environment health level to the remote management platform is as follows: The anomaly detection index is compared with a preset first threshold and a second threshold, wherein the first threshold is less than the second threshold; If the anomaly discrimination index is less than or equal to the first threshold, the health level of the bearing operating environment is determined to be normal. If the anomaly discrimination index is greater than the first threshold and less than the second threshold, the health level of the bearing operating environment is determined to be the warning level. If the anomaly discrimination index is greater than or equal to the second threshold, the health level of the bearing operating environment is determined to be an alarm level. Edge computing nodes encapsulate the anomaly detection index and the corresponding bearing operating environment health level into data packets, which are then sent to the receiving port of the remote management platform via wired or wireless networks.

10. A bearing multi-source sensor data acquisition system based on edge computing, characterized in that, The system is used to implement the bearing multi-source sensing data acquisition method based on edge computing according to any one of claims 1-9, the system comprising: The timing warping module is used by edge computing nodes deployed in the bearing operating environment to align the received bearing environment parameter set and bearing host operating signal set according to timestamps to obtain standardized timing data of the bearing. The convolutional feature extraction module is used to extract local correlation feature maps between environmental parameters and bearing host signals by taking standardized time series data within a continuous time window as input, based on a pre-built deep feature extraction network; The anomaly index mapping module is used to input the local correlation feature map into the recurrent layer, extract the dynamic evolution feature vector along the time axis, and map it into an anomaly discrimination index through edge calculation nodes; The level reporting module is used to determine the bearing operating environment health level based on the anomaly discrimination index, and upload the anomaly discrimination index and bearing environment health level to the remote management platform.

Images