Heterogeneous sensor data-oriented agile access and unified modeling method and device

Abstract

According to the method and device for agile access and unified modeling of heterogeneous sensor data, on the basis of keeping the integrity of original data information, rapid access, structural perception enhanced modeling and channel-level semantic fusion of a multi-source heterogeneous sensor are achieved, and unified and efficient data representation is provided for equipment state evaluation and fault detection. The method comprises the following steps: (1) acquiring sampling original data of each sensor channel; (2) calculating a time step number in a set time window; (3) other low-frequency channels are aligned through interpolation or filling by taking the target length as a reference, and effective data areas and filling areas are marked by using the mask matrix; (4) completing signal value mapping and embedding of position codes; (5) extracting meta-information of each sensor channel; the meta information is coded into a Token; (6) realizing time sequence dynamic capture and multi-channel semantic fusion of information through time-dimensional attention modeling and channel-dimensional attention modeling; and (7) generating unified channel enhanced feature representation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial intelligent equipment testing, and particularly relates to a method for agile access and unified modeling of heterogeneous sensor data and a device for agile access and unified modeling of heterogeneous sensor data. BACKGROUND

[0002] In order to realize state monitoring and health assessment of complex equipment, a large number of sensors with different structures and different sources need to be deployed in the system. These sensors usually have different sampling rates, different unit formats, and inconsistent structures, resulting in strong heterogeneity of original data and great difficulty in integration, and it is difficult to directly access a unified analysis system. The existing solutions are mostly concentrated on the physical access and protocol connection level, and lack of unified representation, automatic semantic modeling, structure context enhancement and cold start support capability of heterogeneous data, which limits the adaptability, response speed and expansion capability of the PHM system. SUMMARY

[0003] In order to overcome the defects of the prior art, the technical problem to be solved by the present application is to provide a method for agile access and unified modeling of heterogeneous sensor data, which can realize rapid access, structure perception enhancement modeling and channel level semantic fusion of multi-source heterogeneous sensors on the basis of maintaining the integrity of original data information, and provide a unified and efficient data representation for equipment state evaluation and fault detection.

[0004] The technical scheme of the present application is that the method for agile access and unified modeling of heterogeneous sensor data comprises the following steps:

[0005] (1) obtaining sampling original data of each sensor channel;

[0006] (2) in a set time window, taking the time step corresponding to the channel with the highest sampling rate as the reference length, calculating the time step n that should be covered in the current time window;

[0007] (3) other low-frequency channels are aligned by interpolation or padding with the target length n as the reference, and the effective data area and the padding area are marked by using a mask matrix;

[0008] (4) mapping the standardized sensor time series data into a tensor form that can be input into the attention mechanism, and completing the embedding operation of signal value mapping and position coding;

[0009] (5) adopting a data protocol analysis mechanism based on a meta-model to extract meta-information of each sensor channel; mapping the information of sensor type, installation position embedding, sampling rate and range into a standardized floating point representation; and then encoding the meta-information into a Token by means of vector-by-vector addition

[0010] Z meta = E pos + E type + E rate + E range

[0011] Concatenate it in front of the input time sequence, form:

[0012]

[0013] Wherein, Z meta is the meta-information vector;

[0014] (6) Introduce a two-stage channel enhancement attention modeling mechanism based on the Transformer structure, realize the time sequence dynamic capture and multi-channel semantic fusion through time dimension attention modeling and channel dimension attention modeling;

[0015] (7) The two-stage output extracted in step (6) is combined to generate a unified channel enhancement feature representation, which is used for subsequent state recognition, fault diagnosis or prediction tasks.

[0016] The application does not depend on channel sample data, and the channel only needs to provide basic meta-information (type, position, sampling rate) to participate in unified attention modeling. The final model output is a structured feature representation, which can be used for downstream classification, fault identification, health scoring, time series prediction tasks, and is suitable for industrial test platforms, remote PHM systems and device self-diagnosis scenarios. On the basis of maintaining the integrity of the original data information, it can realize the rapid access of multi-source heterogeneous sensors, structure perception enhancement modeling and channel level semantic fusion, and provide a unified and efficient data representation for equipment state evaluation and fault detection.

[0017] An apparatus for agile access and unified modeling of heterogeneous sensor data is also provided, and the apparatus comprises:

[0018] An original data acquisition module configured to acquire sampled original data of each sensor channel;

[0019] A step calculation module configured to calculate the time steps n that should be covered in the current time window, with the time step corresponding to the channel with the highest sampling rate as the reference length in the set time window;

[0020] An alignment marking module configured to align other low-frequency channels with the target length n as the reference through interpolation or padding, and mark the effective data area and the padding area using a mask matrix;

[0021] An embedding module configured to map the standardized sensor time series data into a tensor form that can be input into the attention mechanism, complete the embedding operation of signal value mapping and position coding;

[0022] The splicing module is configured to adopt a meta-model-based data protocol parsing mechanism to extract meta-information of each sensor channel; map information of sensor types, installation positions, embedding, sampling rates, and range of quantities into a standardized floating-point representation; and then encode the meta-information into a Token through vector-by-vector addition

[0023] Z meta = E pos + E type + E rate + E range

[0024] Splice it in front of the input time sequence to form:

[0025]

[0026] wherein Z meta is a meta-information vector;

[0027] The fusion module is configured to introduce a two-stage channel enhancement attention modeling mechanism based on a Transformer structure to realize time sequence dynamic capture and multi-channel semantic fusion through time dimension attention modeling and channel dimension attention modeling.

[0028] The enhancement module is configured to jointly generate a unified channel enhancement feature representation from the two-stage output extracted in the fusion module for subsequent state recognition, fault diagnosis, or prediction tasks.

[0029] BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a flowchart of the method for agile access and unified modeling of heterogeneous sensor data according to the present application. DETAILED DESCRIPTION

[0031] In order to make the objectives, technical solutions, and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0032] In order to make the description of the present disclosure more detailed and complete, the following describes the embodiments of the present application and specific examples; however, this is not the only form of implementing or using the specific embodiments of the present application. The embodiments include the features of multiple specific embodiments and the method steps and their order used to construct and operate these specific embodiments. However, other specific embodiments can also be used to achieve the same or equivalent functions and step orders.

[0033] As​Figure 1 As shown, this agile access and unified modeling method for heterogeneous sensor data includes the following steps:

[0034] (1) Obtain the original data of each sensor channel sampling;

[0035] (2) In the set time window, take the time step corresponding to the channel with the highest sampling rate as the reference length, and calculate the time step n that should be covered in the current time window;

[0036] (3) Other low-frequency channels are aligned by interpolation or padding with the target length n as the reference, and the mask matrix is used to mark the effective data area and the padding area;

[0037] (4) Map the normalized sensor time series data into a tensor form that can be input into the attention mechanism, complete the signal value mapping and position encoding embedding operation;

[0038] (5) Adopt a meta-model-based data protocol parsing mechanism to extract the meta-information of each sensor channel;

[0039] Map the information of sensor type, installation position, sampling rate, and range to a standardized floating point representation, and then encode the meta-information into a Token by vector-wise addition

[0040] Z meta = E pos + E type + E rate + E range

[0041] Concatenate it in front of the input time series sequence to form:

[0042]

[0043] Where, Z meta is the meta-information vector;

[0044] (6) Introduce a two-stage channel enhancement attention modeling mechanism based on the Transformer structure, realize the time sequence dynamic capture and multi-channel semantic fusion of information through time dimension attention modeling and channel dimension attention modeling;

[0045] (7) Jointly generate a unified channel enhancement feature representation from the two-stage output extracted in step (6),

[0046] Used for subsequent state recognition, fault diagnosis or prediction tasks.

[0047] The application does not depend on channel sample data, and the channel only needs to provide basic meta information (type, position, sampling rate) to participate in unified attention modeling. The final model output is a structured feature representation, which can be used for downstream classification, fault identification, health scoring, time series prediction tasks, and is suitable for industrial test platforms, remote PHM systems and device self-diagnosis scenarios. On the basis of maintaining the integrity of the original data information, the application can realize the rapid access of multi-source heterogeneous sensors, structure perception enhanced modeling and channel-level semantic fusion, and provide unified and efficient data representation for equipment state evaluation and fault detection.

[0048] Preferably, in the step (4), the signal value is mapped as follows:

[0049] At each time step t, the input is a multi-channel vector x t ∈R C , C is the number of channels, and a linear mapping layer or a 1D convolution layer shared by one channel is used:

[0050] E t = MLP(x t ) ∈R d

[0051] The entire sequence becomes

[0052] X embed = [E1, E2,..., E T ] ∈R T×d .

[0053] Preferably, in the step (4), the step of position encoding introduces a position vector PE t ∈R d for each time step, which is directly added to E t , and the overall input is:

[0054] Z t = E t + PE t ∈R d

[0055] The finally formed input tensor has the shape:

[0056] Z ∈R T×d

[0057] Where, T represents the length of the time sequence contained in each sample, and d represents the feature vector dimension mapped from each time step.

[0058] Preferably, in the step (5), for the discrete numerical information of sensor type and installation position embedding, a word vector table is created to map into an embedding vector.

[0059] Preferably, in the step (5), for numerical type information, the normalized processing is mapped to a vector space through MLP.

[0060] Preferably, in the step (6), the time dimension attention modeling is modeled through a multi-head self-attention mechanism, and the dynamic characteristics in the time dimension inside each channel are modeled to extract the time sequence dependent mode and global context information.

[0061] Preferably, in the step (6), the channel dimension attention modeling module splices the channel representation and its meta information vector to form a channel level feature, and then inputs the channel level feature into a channel dimension attention network to give higher attention weight to the key channel while suppressing the weakly related or missing channel.

[0062] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment methods can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium. When the program is executed, each step of the above-mentioned embodiment method is included, and the storage medium can be ROM / RAM, a magnetic disc, an optical disc, a memory card, etc. Therefore, corresponding to the method of the present application, the present application also simultaneously includes a device for agile access and unified modeling of heterogeneous sensor data. The device is usually represented in the form of a functional module corresponding to each step of the method. The device includes:

[0063] An original data acquisition module configured to acquire sampling original data of each sensor channel;

[0064] A step number calculation module configured to calculate the time step number n to be covered in the current time window based on the time step number corresponding to the channel with the highest sampling rate as the reference length in the set time window;

[0065] An alignment marking module configured to align other low-frequency channels with the target length n as the reference through interpolation or padding, and mark the effective data area and the padding area by using a mask matrix;

[0066] An embedding module configured to map the standardized sensor time sequence data into a tensor form that can be input into an attention mechanism, and complete the embedding operation of signal value mapping and position coding;

[0067] A splicing module configured to use a meta-model-based data protocol parsing mechanism to extract the meta information of each sensor channel; map the information of sensor type, installation position embedding, sampling rate, and range of the range into a standardized floating point representation; and then encode the meta information into a Token by means of vector-by-vector addition

[0068] Z meta = E pos + E type+E rate +E range

[0069] Concatenate it in front of the input time sequence, form:

[0070]

[0071] Where, Z meta is the meta-information vector;

[0072] The fusion module is configured to introduce a two-stage channel enhancement attention modeling mechanism based on the Transformer structure, and realize the time sequence dynamic capture and multi-channel semantic fusion of information through time dimension attention modeling and channel dimension attention modeling;

[0073] The enhancement module is configured to jointly generate a unified channel enhancement feature representation from the two-stage output extracted in the fusion module, for subsequent state recognition, fault diagnosis or prediction tasks.

[0074] Preferably, in the embedding module, the signal value is mapped as follows:

[0075] At each time step t, the input is a multi-channel vector x t ∈R C , C is the number of channels, and a channel-shared linear mapping layer or a 1D convolution layer is used:

[0076] E t = MLP(x t ) ∈R d

[0077] The entire sequence becomes

[0078] X embed = [E1, E2,..., E T ] ∈R T×d ;

[0079] The step of position encoding introduces a position vector PE t ∈R d for each time step, which is directly added to E t , and the overall input is:

[0080] Z t = E t + PE t ∈R d

[0081] The final input tensor formed has the shape:

[0082] Z ∈R T×d

[0083] Where, T represents the length of time series contained in each sample, and d represents the dimension of feature vector mapped from each time step.

[0084] Preferably, in the splicing module, for the discrete numerical information embedded in the sensor type and installation position, a word vector table is created to map into an embedding vector; for numerical information, after normalization, it is mapped into a vector space through an MLP.

[0085] In the fusion module, the time dimension attention modeling is modeled through a multi-head self-attention mechanism to model the dynamic characteristics of each channel in the time dimension and extract the time sequence dependency pattern and global context information; the channel dimension attention modeling module splices the channel representation and its meta information vector to form a channel level feature, and then inputs it into a channel dimension attention network to give higher attention weight to the key channel while suppressing the weakly related or missing channels.

[0086] The application will be described in detail below with reference to the accompanying drawings and examples.

[0087] The application proposes an agile access and unified modeling method for heterogeneous sensor data, as shown in the figure, the method comprises the following steps: Figure 1

[0088] Step one: obtain the original sampling data of each sensor channel;

[0089] Step two: in the set time window, take the time step number corresponding to the channel with the highest sampling rate as the reference length, calculate the time step number n that should be covered in the current time window

[0090] Step three: other low-frequency channels are aligned through interpolation or padding. If the original time series is less than the target length n, zero padding is performed, and a mask matrix is used to mark the effective data area and the padding area to ensure data integrity and unified input structure of the model;

[0091] Step four: map the normalized sensor time series data into a tensor form that can be input into the attention mechanism (such as Transformer), i.e. the embedding operation before the encoder.

[0092] 1) Signal value mapping

[0093] At each time step t, the input is a multi-channel vector x t ∈R C (C is the number of channels)

[0094] A channel-shared linear mapping layer or a 1D convolution layer is used:

[0095] E t = MLP(x t ) ∈R d ​

[0096] The entire sequence becomes

[0097] X embed = [E1, E2,..., E T ] ∈ R T×d

[0098] If the channel semantics differ greatly, an independent linear layer or small MLP can also be set for each channel before splicing.

[0099] 2) Position encoding

[0100] A position vector PE t ∈ R d is introduced for each time step and directly added to E t ; The form of sine and cosine position encoding is adopted, that is:

[0101] PE (t,2i) = sin(t / 10000 2i / d )

[0102] PE (t,2i+1) = cos(t / 10000 2i / d )

[0103] The overall input is:

[0104] Z t = E t + PE t ∈ R d

[0105] The final input tensor shape is:

[0106] Z ∈ R T×d

[0107] Where T represents the length of the time series contained in each sample, and d represents the feature vector dimension mapped by each time step. This tensor can be directly used as the input of the Transformer encoder.

[0108] Step five: Adopt the meta-model-based data protocol parsing mechanism to extract the meta-information of each sensor channel. Map the information such as sensor type, installation position embedding, sampling rate, and range to a standardized floating-point representation. The specific operation steps are as follows:

[0109] 1) For discrete numerical information such as sensor type and installation position embedding, create a word vector table to map to an embedding vector. For example, the installation position embedding Sensor type category embedding

[0110] 2) For numerical information such as sampling rate and range, etc., normalization is performed and then mapped to a vector space by MLP, such as mapping sampling rate and range to

[0111] 3) The above meta information is encoded into a special Token by vector-by-vector addition, that is,

[0112] Z meta = E pos + E type + E rate + E range

[0113] When spliced in front of the input time sequence of the Transformer, it forms:

[0114]

[0115] Where, Z meta is the meta information vector, which can participate in attention interaction.

[0116] Step six: The model introduces a two-stage channel-enhanced attention modeling mechanism based on the Transformer structure, the core goal of which is to model the time dependence and channel synergy on the unified input representation vector Z, to realize the dynamic capture of information and multi-channel semantic fusion. The mechanism includes the following steps:

[0117] 1) Time dimension attention modeling, used to model the time series of each channel independently;

[0118] Input: Time series tensor of each channel, with time position coding and frequency modulation added, combined into input vector Where C is the number of channels, T is the number of time steps, and d is the embedding dimension;

[0119] Processing: Through the multi-head self-attention mechanism, the dynamic features of each channel in the time dimension are modeled, and the time sequence dependence pattern and global context information are extracted;

[0120] Output: Time dimension feature compression vector of each channel Forming the first stage output set

[0121] 2) Channel dimension attention modeling module, used to model the interaction between channels at each time step;

[0122] Input: Channel features of the first stage output Channel meta information embedding, including structural position, component type embedding, sensor category encoding, etc. Z meta ;

[0123] Processing: All channel representations are concatenated with their meta-information vectors to form channel-level features; input into the channel-wise attention network; give higher attention weight to key channels, while suppressing weakly related or missing channels;

[0124] Output: Aggregated global channel semantic representation Z C , represents the fusion state of the entire channel set under the current window.

[0125] Step seven: The two-stage output jointly generates a unified channel enhancement feature representation, which is used for subsequent state recognition, fault diagnosis or prediction tasks.

[0126] The above is only a preferred embodiment of the present application, not any form of limitation on the present application, any simple modification, equivalent change and modification of the above embodiment according to the technical essence of the present application are still within the protection scope of the technical solution of the present application.

Claims

1. A method for agile access and unified modeling of heterogeneous sensor data, characterized by: It includes the following steps: (1) Obtain the raw sampling data of each sensor channel; (2) Within the set time window, take the time step corresponding to the channel with the highest sampling rate as the base length and calculate the number of time steps n that should be covered in the current time window; (3) Other low-frequency channels are aligned by interpolation or padding based on the target length n, and the effective data area and padding area are marked by a mask matrix. (4) The standardized sensor time series data is mapped into a tensor form that can be input into the attention mechanism to complete the embedding operation of signal value mapping and position encoding; (5) A data protocol parsing mechanism based on a meta-model is adopted to extract the meta-information of each sensor channel; the information of sensor type, installation location embedding, sampling rate, and measurement range is uniformly mapped into a standardized floating-point representation; and then the meta-information is encoded into a token by vector-wise summation. Z meta =And pos +E type +E rate +E range It is appended to the beginning of the input time series to form: Among them, Z meta It is a meta-information vector; (6) Introduce a two-stage channel-enhanced attention modeling mechanism based on Transformer structure, and realize the temporal dynamic capture of information and multi-channel semantic fusion through time-dimensional attention modeling and channel-dimensional attention modeling; (7) The two-stage outputs extracted in step (6) are combined to generate a unified channel enhancement feature representation for subsequent state recognition, fault diagnosis or prediction tasks.

2. The method for agile access and unified modeling of heterogeneous sensor data according to claim 1, characterized in that: In step (4), the signal value is mapped as follows: At each time step t, the input is a multi-channel vector x. t ∈R C C is the number of channels, using a linear mapping layer or a 1D convolutional layer with one shared channel: E t <MLP(x t )∈R d The entire sequence becomes X embed =[E1,E2,...,E T ]∈R T×d .

3. The method for agile access and unified modeling of heterogeneous sensor data according to claim 2, characterized in that: In step (4), the position encoding step introduces a position vector PE for each time step. t ∈R d Add directly to E t Above, enter the entire text: Z t =E t +PE t ∈R d The final input tensor shape is: Z∈R T×d Where T represents the length of the time series contained in each sample, and d represents the dimension of the feature vector mapped to each time step.

4. The method for agile access and unified modeling of heterogeneous sensor data according to claim 3, characterized in that: In step (5), for the discrete numerical information embedded in the sensor type and installation location, a word vector table is created and mapped to the embedded vector.

5. The method for agile access and unified modeling of heterogeneous sensor data according to claim 4, characterized in that: In step (5), for numerical information, after normalization, it is mapped to the vector space through MLP.

6. The method for agile access and unified modeling of heterogeneous sensor data according to claim 5, characterized in that: In step (6), the temporal attention modeling uses a multi-head self-attention mechanism to model the dynamic features of each channel in the temporal dimension and extract the temporal dependency pattern and global context information.

7. The method for agile access and unified modeling of heterogeneous sensor data according to claim 6, characterized in that: In step (6), the channel-dimensional attention modeling module concatenates the channel representation with its meta-information vector to form channel-level features, which are then input into the channel-dimensional attention network to give higher attention weights to key channels while suppressing weakly correlated or missing channels.

8. A device for agile access and unified modeling of heterogeneous sensor data, characterized in that: The device includes: The raw data acquisition module is configured to acquire raw data sampled from each sensor channel. The step count calculation module is configured to calculate the number of time steps n that should be covered within the current time window, using the time step count corresponding to the channel with the highest sampling rate as the baseline length. The alignment marking module is configured to align other low-frequency channels with a target length n as a reference, through interpolation or padding, and to mark the valid data area and the padding area using a mask matrix; The embedding module is configured to map the standardized sensor time-series data into a tensor form that can be input into the attention mechanism, completing the embedding operations of signal value mapping and location encoding. The stitching module is configured to use a data protocol parsing mechanism based on a meta-model to extract the meta-information of each sensor channel; uniformly map the sensor type, installation location embedding, sampling rate, and measurement range information into a standardized floating-point representation; and then encode the meta-information into a token through vector-wise summation. Z meta =And pos +E type +E rate +E range It is appended to the beginning of the input time series to form: Among them, Z meta It is a meta-information vector; The fusion module is configured to introduce a two-stage channel-enhanced attention modeling mechanism based on the Transformer structure, which realizes the temporal dynamic capture of information and multi-channel semantic fusion through temporal attention modeling and channel-dimensional attention modeling. The enhancement module is configured to jointly generate a unified channel enhancement feature representation from the two-stage outputs extracted by the fusion module for subsequent state recognition, fault diagnosis, or prediction tasks.

9. The apparatus for agile access and unified modeling of heterogeneous sensor data according to claim 8, characterized in that: In the embedded module, the signal value mapping is as follows: At each time step t, the input is a multi-channel vector x. t ∈R C C is the number of channels, using a linear mapping layer or a 1D convolutional layer with one shared channel: E t <MLP(x t )∈R d The entire sequence becomes X embed =[E1,E2,...,E T ]∈R T×d ; The position encoding step introduces a position vector PE for each time step. t ∈R d Add directly to E t Above, enter the entire text: Z t =E t +PE t ∈R d The final input tensor shape is: Z∈R T×d Where T represents the length of the time series contained in each sample, and d represents the dimension of the feature vector mapped to each time step.

10. The apparatus for agile access and unified modeling of heterogeneous sensor data according to claim 9, characterized in that: In the splicing module, for discrete numerical information embedded by sensor type and installation location, a word vector table is created and mapped to embedded vectors; for numerical information, after normalization processing, it is mapped to vector space through MLP. In the fusion module, the temporal attention modeling uses a multi-head self-attention mechanism to model the dynamic features of each channel in the temporal dimension and extract temporal dependency patterns and global context information. The channel-dimensional attention modeling module concatenates the channel representation with its meta-information vector to form channel-level features, which are then input into the channel-dimensional attention network to give higher attention weights to key channels while suppressing weakly correlated or missing channels.

Images