A multi-sensor multi-component health prediction method and apparatus

Abstract

The application provides a multi-sensor multi-component health prediction method and device. The method provided by the application comprises: collecting multi-source sensor data of different components in a device, time aligning and preprocessing the multi-source sensor data, and constructing a multi-channel data matrix with a unified spatial structure; for each sensor channel in the multi-channel data matrix, respectively fusing a time sequence feature, a channel identifier feature and a physical component attribute feature to construct a corresponding independent channel feature vector; taking the independent channel feature vectors of the sensor channels as inputs of a reverse Transformer architecture, performing self-attention calculation on the sensor channels as feature units, introducing a physical constraint bias and a mask bias, and outputting a feature matrix encoding spatial coupling relationships between different sensors; aggregating the feature matrix, extracting a comprehensive shared representation of the device, and outputting health prediction results for multiple components in the device in parallel based on the comprehensive shared representation.

Description

Technical Field

[0001] This application relates to the field of health prediction technology, and in particular to a multi-sensor, multi-component health prediction method and apparatus. Background Technology

[0002] As a key core device in the tunnel ventilation system, the operation status of subway axial flow fans directly affects the safe operation of trains and the quality of the tunnel environment. Because they operate under high load and continuous operation for a long time and are composed of multiple components working together, single-dimensional monitoring data is difficult to fully reflect the overall health trend. Therefore, dynamic health prediction that covers multi-sensor data and multi-component collaborative modeling is of great significance for ensuring the safe and efficient operation and maintenance of rail transit.

[0003] Currently, health prediction for such equipment mainly relies on the following approaches: first, threshold or model-based judgment based on single sensor data such as vibration or temperature; second, modeling single-channel time series data using traditional machine learning or deep learning algorithms such as convolutional neural networks and long short-term memory networks; and third, simply stitching together or independently modeling multi-source sensor data before fusing the results at the output end. However, these methods have significant shortcomings in actual industrial scenarios. They struggle to effectively model the deep spatial coupling relationships between different sensors, lack cross-component collaborative analysis capabilities, and are poorly adaptable to changes in the number of sensors. Once signal damage or loss occurs, traditional networks will directly fail. Furthermore, purely data-driven methods do not incorporate the physical mechanisms of the equipment, resulting in a lack of interpretability in the model prediction process.

[0004] Therefore, there is an urgent need for a method to improve the accuracy of multi-sensor cross-component health prediction and the fault tolerance robustness in dynamic missing scenarios in complex industrial environments. Summary of the Invention

[0005] In view of this, this application provides a multi-sensor, multi-component health prediction method and apparatus to improve the accuracy of multi-sensor, cross-component health prediction in complex industrial environments and the fault tolerance and robustness in dynamic missing scenarios.

[0006] Specifically, this application is implemented through the following technical solution:

[0007] The first aspect of this application provides a multi-sensor, multi-component health prediction method, the method comprising:

[0008] Collect multi-source sensor data from different components in the device, perform time alignment and preprocessing on the multi-source sensor data, and construct a multi-channel data matrix with a unified spatial structure;

[0009] For each sensor channel in the multi-channel data matrix, time-series features, channel identification features, and physical component attribute features are fused to construct a corresponding independent channel feature vector;

[0010] The independent channel feature vectors of each sensor channel are used as input to the inverse Transformer architecture. Self-attention calculation is performed with the sensor channel as the feature unit. In the process of self-attention calculation, physical constraint biases that characterize the mechanism relationship between channels and mask biases that characterize the dynamic missing state of channels are introduced. The output is a feature matrix that encodes the spatial coupling relationship between different sensors.

[0011] The feature matrix is ​​aggregated, a comprehensive shared representation of the device is extracted, and health prediction results for multiple components in the device are output in parallel based on the comprehensive shared representation.

[0012] A second aspect of this application provides a multi-sensor, multi-component health prediction device, the device comprising a construction module and an output module;

[0013] The construction module is used to collect multi-source sensor data from different components in the device, perform time alignment and preprocessing on the multi-source sensor data, and construct a multi-channel data matrix with a unified spatial structure.

[0014] The construction module is also used to fuse time-series features, channel identification features, and physical component attribute features for each sensor channel in the multi-channel data matrix to construct a corresponding independent channel feature vector;

[0015] The output module is used to take the independent channel feature vectors of each sensor channel as input to the inverse Transformer architecture, and perform self-attention calculation with the sensor channel as the feature unit. In the self-attention calculation process, physical constraint biases that characterize the mechanism relationship between channels and mask biases that characterize the dynamic missing state of channels are introduced, and the output is a feature matrix that encodes the spatial coupling relationship between different sensors.

[0016] The output module is also used to aggregate the feature matrix, extract the comprehensive shared representation of the device, and output health prediction results for multiple components in the device in parallel based on the comprehensive shared representation.

[0017] The multi-sensor, multi-component health prediction method and apparatus provided in this application firstly collect multi-source sensor data from different components and perform time-aligned preprocessing to construct a multi-channel data matrix. This solves the pain points of inconsistent sampling rates and disconnected time-domain references among heterogeneous sensors, and achieves standardized spatial alignment of global multi-dimensional information. Subsequently, independent channel feature vectors are constructed by fusing time series, channel identifiers, and physical component attribute features. Prior spatial topology and device structural attributes are explicitly injected at the token level, giving the initial representation strong spatiotemporal discriminability. Independent vectors are input into the inverse Transformer architecture, and self-attention computation is performed using channels as feature units. Physical constraint biases and mask biases are introduced to output a spatially coupled feature matrix. On one hand, the inverse architecture transforms the attention focus from traditional time series to cross-channel spatial correlation mining. On the other hand, physical constraint biases characterizing the mechanistic correlations between channels are introduced, rigidly embedding structural topology mechanisms into matrix operations, suppressing spurious correlations caused by noise, and achieving a proactive dual gain in interpretability and accuracy. Simultaneously, mask biases characterizing the dynamic missing states of channels are introduced, providing an online dynamic fault tolerance mechanism for field anomalies such as hardware damage or communication interruptions, isolating and suppressing interference from failed channels in real time, and endowing the system with extremely strong structural noise resistance. Finally, a comprehensive shared representation is extracted by aggregating the feature matrix, and the health prediction results of multiple components are output in parallel. Feature sharing is maximized at the underlying level, significantly reducing the massive computational overhead of repeated network construction and effectively solving the bottleneck of collaborative evolution under strong coupling of multiple components, achieving a highly robust and highly consistent system-level intelligent operation and maintenance decision-making effect. Attached Figure Description

[0018] Figure 1 A flowchart of the multi-sensor, multi-component health prediction method provided in Embodiment 1 of this application;

[0019] Figure 2 This is a schematic diagram of the structure of the multi-sensor, multi-component health prediction device provided in Embodiment 2 of this application. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0021] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0022] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0023] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0024] Figure 1 This is a flowchart of the multi-sensor, multi-component health prediction method provided in Embodiment 1 of this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0025] S101. Collect multi-source sensor data from different components in the device, perform time alignment and preprocessing on the multi-source sensor data, and construct a multi-channel data matrix with a unified spatial structure.

[0026] Specifically, in this application, "equipment" specifically refers to a subway axial flow fan, which is a key core component of the tunnel ventilation system. Multi-source sensor data refers to diverse and heterogeneous raw signals collected from different core components of the aforementioned equipment, used to characterize the real-time operating status of the equipment from multiple physical dimensions. In this application, multi-source sensor data includes voltage signals, current signals, frequency signals, temperature signals, vibration signals, and noise signals. In their raw state, they have different sampling frequencies, amplitude dimensions, and dispersed data structures. A multi-channel data matrix refers to a regular two-dimensional numerical matrix formed by combining multi-source sensor signals that were originally misaligned in time and dispersed in space, according to a preset channel order, after processing.

[0027] In practice, a subway axial flow fan is used as the device to be monitored, and several different core components within the device that require status monitoring are identified. Corresponding types of sensor hardware are then strategically deployed and installed on each of these core components. Through these deployed sensors, various heterogeneous raw data from different components are synchronously read and collected.

[0028] Furthermore, according to a preset channel order, the multi-source sensor data are arranged spatially to construct an initial multi-channel data matrix. For the initial multi-channel data matrix, the signal of the first channel is downsampled, and the signal of the second channel is interpolated. The data of all sensor channels are aligned to a unified time window for resampling. The sampling rate of the first channel is greater than that of the second channel. The statistical mean and standard deviation of each sensor channel within the time window after resampling are calculated respectively. The resampled data value of each channel is subtracted from the corresponding statistical mean and then divided by the corresponding standard deviation to output a standardized multi-channel data matrix. A corresponding mask vector is constructed according to the data integrity status of each sensor channel at the current time, and the standardized multi-channel data matrix and the mask vector are output synchronously.

[0029] In practical implementation, a fixed channel index mapping table is predefined in system memory (e.g., channel 0 - voltage, channel 1 - current, channel 2 - frequency, channel 3 - temperature, channel 4 - vibration, channel 5 - noise). Sensor data streams deployed on different components of the subway axial flow fan are read simultaneously via multi-threading or a parallel bus. Heterogeneous data collected within the same relative time period and scattered in different data buffers are forcibly aligned and arranged according to a preset index order, forming a two-dimensional array or tensor as the initial multi-channel raw data matrix. Metadata for each channel is read, and the sampling frequency is automatically identified. High-frequency channels (such as high-frequency vibration and noise signals, designated as the first channel) and low-frequency channels (such as low-frequency temperature and static voltage signals, designated as the second channel) are classified. High-frequency signals are first filtered using a low-pass anti-aliasing filter (such as a Butterworth filter) to prevent spectral aliasing; then, they are decimated or downsampled according to integer step sizes to reduce the sampling rate to the system's set reference target frequency. For low-frequency signals, mathematical interpolation algorithms (such as linear interpolation, cubic spline interpolation, or nearest neighbor interpolation) are used to calculate and fill in missing time-series data points among known sparse data points, increasing the sampling rate to the same reference target frequency. A fixed-length time-domain sliding window (e.g., a time window containing T time steps) is set, and all resampled, fully aligned channel signals are extracted to form a regularized time-series matrix of dimension T×C (time step size × number of channels). The time-series matrix is ​​traversed along its time axis (row direction), and for each channel (column direction), the arithmetic mean and standard deviation of all values ​​within its window are calculated independently. For each element in the entire matrix, the original time-series value of the element is subtracted from the average value of its corresponding channel, and then divided by the standard deviation of that channel to obtain a standardized multi-channel data matrix. Furthermore, within the current time window, the hardware connection and data reception status of each sensor channel are checked in real time. A one-dimensional Boolean or integer vector with a length equal to the total number of channels C is allocated in memory. The system iterates through the inspection results. If a channel's signal is complete and normal within the current window, a standard flag (such as 0 or True) is assigned to the corresponding position in the mask vector. If data loss due to sensor damage, communication interruption, or disconnection is detected in a channel, a defect flag (such as 1 or False) is assigned to the corresponding position. Using structure or tuple data encapsulation techniques, a standardized multi-channel data matrix of T×C dimensions is forcibly bound to a dynamic mask vector of length C within the same data packet, serving as a unified input source.

[0030] For example, in one embodiment, assuming there are a total of C sensor channels, the multi-source sensor data can be represented as:

[0031] ;

[0032] in, Data from multiple sensor sources; For the first The sensor channel in the first Sensor values ​​collected at each time step; This represents the total number of sensor channels. For the first The total number of raw sampling points for each sensor channel during the current acquisition cycle.

[0033] The resampling process can be represented as:

[0034] ;

[0035] in, For the resampling of the first The sensor channel in the first Sensor values ​​collected at each time step; For the first The sensor channel in the first Sensor values ​​collected at each time step; To ensure uniform alignment and standard time window length after truncation.

[0036] Calculate the statistical mean and standard deviation of each sensor channel within the time window:

[0037] ;

[0038] ;

[0039] in, This is the statistical mean; Standard deviation; For the resampling of the first The sensor channel in the first Sensor values ​​collected at each time step; To ensure uniform alignment and standard time window length after truncation.

[0040] Output the corresponding elements in the normalized multichannel data matrix:

[0041] ;

[0042] in, To standardize the corresponding elements in the multi-channel data matrix; This is the statistical mean; Standard deviation; For the resampling of the first The sensor channel in the first Sensor values ​​collected at each time step.

[0043] A standardized multichannel data matrix can be represented as:

[0044] ;

[0045] in, To standardize the multi-channel data matrix; This represents the total number of sensor channels. To ensure uniform alignment and standard time window length after truncation.

[0046] The mask vector can be represented as:

[0047] ;

[0048] in, It is a mask vector; For the corresponding number in the mask vector The state mask element for each sensor channel.

[0049] S102. For each sensor channel in the multi-channel data matrix, fuse the timing features, channel identification features, and physical component attribute features to construct a corresponding independent channel feature vector.

[0050] Specifically, temporal features refer to the dynamic changes of the sensor over a given period. Channel identification features identify each sensor, explicitly telling the model "I am a vibration sensor" or "I am a temperature sensor," preventing semantic confusion caused by heterogeneous data from multiple sources after being fed into the attention mechanism. Physical component attribute features refer to the mechanistic label indicating which key part of the sensor is installed. Independent channel feature vectors refer to the final generated all-around feature token representing that sensor channel, which is the final product after fusing the above three features in the feature space.

[0051] In specific implementation, the standardized time series data of each sensor channel are sequentially subjected to one-dimensional convolution, activation, and pooling to extract time feature codes containing time-domain statistical features and frequency-domain features; embedding mapping is performed according to the channel index of each sensor channel to generate channel codes to distinguish the spatial location of different channels; embedding mapping is performed according to the type of component monitored by each sensor channel to generate physical attribute codes to characterize the specificity of different physical components; the time feature codes, channel codes, and physical attribute codes are summed, element-aligned, and fused to output independent channel feature vectors.

[0052] Specifically, the standardized multi-channel data matrix is ​​split into standardized time-series data for individual sensor channels. The time-series data for each sensor channel is input into a one-dimensional convolutional layer (1D CNN), where sliding computation is performed using a convolutional kernel with a set stride and kernel size to initially extract local temporal features. The output of the convolution is then fed into an activation layer (such as ReLU or GELU functions) for nonlinear mapping, capturing nonlinear frequency and temporal features. The activated data is then fed into a pooling layer (such as Max Pooling or Average Pooling) for downsampling, compressing the feature dimension and extracting global temporal statistical and frequency features, outputting the final temporal feature code. The spatial arrangement of the current sensor channel in the system is hard-coded (e.g., unique index labels such as channel 0, channel 1, etc.). This channel index is input into a learnable channel embedding layer for lookup mapping, converting it into a high-dimensional vector with the same dimension as the temporal feature code, outputting the channel code used to distinguish the spatial locations of different channels. Identify the specific monitoring component type (e.g., impeller, bearing, stator, etc.) of the subway axial flow fan corresponding to the current sensor channel. Input this component type into the Component Embedding Layer for mapping, converting it into a high-dimensional vector of the same dimension, and outputting physical attribute codes to characterize the specificity of different physical components. Align the extracted time feature codes, channel codes, and physical attribute codes to a unified dimension. In the feature space, add the elements at the same position corresponding to these three codes element by element. Encapsulate the composite high-dimensional tensor obtained after addition and fusion as an independent channel feature vector.

[0053] For example, in one embodiment, the time feature encoding can be represented as:

[0054] ;

[0055] in, Encode the time feature vector; This refers to the corresponding elements in the standardized multi-channel data matrix.

[0056] Channel encoding can be represented as:

[0057] ;

[0058] in, This is the channel encoding vector; Embed operators for pre-defined learnable channels; For sensor channels.

[0059] Physical attribute encoding can be represented as:

[0060] ;

[0061] in, Encode vectors for physical properties; Embed operators for pre-defined learnable physical properties; For the first The label for the type of physical component monitored by each sensor channel.

[0062] Independent channel feature vectors can be represented as:

[0063] ;

[0064] in, These are the feature vectors of independent channels; Encode the time feature vector; This is the channel encoding vector; Encode vectors for physical properties.

[0065] The feature matrix of independent channels can be represented as:

[0066] ;

[0067] in, The feature matrix is ​​for each independent channel; These are the feature vectors of independent channels; This represents the total number of sensor channels. The preset high-dimensional feature mapping dimension.

[0068] S103. The independent channel feature vectors of each sensor channel are used as input to the inverse Transformer architecture. Self-attention calculation is performed with the sensor channel as the feature unit. In the process of self-attention calculation, physical constraint biases that characterize the mechanism relationship between channels and mask biases that characterize the dynamic missing state of channels are introduced. The output is a feature matrix that encodes the spatial coupling relationship between different sensors.

[0069] Specifically, traditional Transformers, when processing multivariate time-series data, typically package all sensor values ​​at the same time step into a single token (feature unit), allowing the attention mechanism to calculate the temporal correlation between time steps. This can easily lead to semantic confusion within the token due to sensors with different dimensions, and makes it difficult to capture spatial interactions across sensors. This application adopts an inverse architecture, treating the complete normalized time series of a single sensor channel as an independent token. In this inverse architecture, the self-attention mechanism no longer calculates the causal sequence of time, but rather the spatial correlation between different sensor channels. For example, it allows the model to directly calculate the mutual influence weights between the "main bearing temperature channel" and the "main bearing vibration channel," thus being inherently suitable for uncovering the spatial coupling relationships between multi-source heterogeneous sensors.

[0070] Physical constraint bias refers to the association bias matrix introduced in channel self-attention computation, based on the actual mechanical topology of the equipment and pre-defined industrial mechanisms. It is used to constrain the association computation of the purely data-driven model, ensuring that the learned features conform to real physical laws and prior common sense. Mask bias refers to the control matrix constructed using dynamic state masks and applied to the self-attention mechanism. It can automatically identify faulty channels with missing or corrupted data and forcibly reset their attention weights to zero, thereby cutting off the interference of abnormal data on global collaborative analysis. Feature matrix refers to the high-order global state feature tensor finally output by the self-attention mechanism. Each channel vector in this matrix is ​​no longer isolated but deeply integrated with the information of all other channels, and can fully characterize the spatial coupling and collaborative evolution relationship between different sensors.

[0071] Optionally, the physical constraint matrix constructed by the physical constraint bias includes: constructing a square matrix whose number of rows and columns are equal to the total number of sensor channels as the physical constraint matrix; analyzing whether there is a physical coupling relationship between any two sensor channels based on the knowledge of the device's operating mechanism; if the corresponding first sensor channel and second sensor channel are related in terms of physical structure or operating mechanism, assigning a preset positive characteristic correlation constant to the intersection position of the first sensor channel and the second sensor channel in the physical constraint matrix according to the strength of the physical correlation between the first sensor channel and the second sensor channel; the positive characteristic correlation constant is used to amplify the attention weight between the first sensor channel and the second sensor channel; if there is no physical mechanism relationship between the corresponding first sensor channel and the second sensor channel, assigning a zero value to the corresponding intersection position in the physical constraint matrix.

[0072] Specifically, obtain the total number of sensor channels C in the current system. Create a two-dimensional square matrix (C×C) in memory with C rows and columns, and define it as the physical constraint matrix. Import pre-defined operating mechanism knowledge such as the mechanical topology, aerodynamics, or heat transfer of the subway axial flow fan. Pair and check any two sensors corresponding to rows (representing the first sensor channel) and columns (representing the second sensor channel) in the physical constraint matrix to analyze whether they have coupling relationships in the actual physical environment. If it is determined that the first sensor channel and the second sensor channel are related in terms of physical structure (e.g., mounted on the same bearing housing) or operating mechanism (e.g., dynamic transmission between temperature and vibration), further evaluate the strength of their physical relationship. Based on the evaluated relationship strength, assign a pre-defined positive feature correlation constant (value greater than 0) to the corresponding row and column intersection coordinate position in the physical constraint matrix to directly amplify the attention weight between the two channels in the subsequent attention mechanism calculation. The positive feature correlation constant is positively correlated with the strength of the physical relationship. If it is determined that the first sensor channel and the second sensor channel are completely independent in terms of physical structure or operating mechanism and have no coupling relationship. The corresponding row and column intersection coordinates in the physical constraint matrix are assigned zero values, and no mechanistic amplification is applied to the attention weights of the two in subsequent calculations.

[0073] For example, in one embodiment, the initially constructed physical constraint matrix can be represented as:

[0074] ;

[0075] in, This is the initially constructed physical constraint matrix; This represents the total number of sensor channels.

[0076] The assignment formula for the corresponding elements in the physical constraint matrix can be expressed as:

[0077] ;

[0078] in, These are the corresponding elements in the physical constraint matrix; It is a positive characteristic correlation constant.

[0079] The updated attention weights can be represented as:

[0080] ;

[0081] in, The updated attention weights; These are the original attention weights; This refers to the corresponding element in the physical constraint matrix.

[0082] Optionally, the mask matrix constructed by the mask bias includes: constructing a square matrix whose number of rows and columns are equal to the total number of sensor channels as the mask matrix; obtaining the status identifier value corresponding to each sensor channel according to the available status of each sensor channel at the current time; if the corresponding first sensor channel and the second sensor channel are both in a normal state, assigning a zero value to the intersection position of the first sensor channel and the second sensor channel in the mask matrix; if at least one of the corresponding first sensor channel and / or second sensor channel is in a data missing state, assigning a negative infinity value to the corresponding intersection position in the mask matrix.

[0083] Specifically, obtain the total number C of sensor channels in the current system. Create a two-dimensional square matrix (C×C) in memory with C rows and columns, and define it as a mask matrix. Real-time detection or extraction of the operational and communication availability status of each sensor channel in the device at the current moment. Based on the detection results, obtain the status identifier value corresponding to each sensor channel (to characterize whether the channel is in a normal operating state or a data missing state). Check the status of any two sensors corresponding to rows (representing the first sensor channel) and columns (representing the second sensor channel) in the mask matrix one by one. If it is determined that both the corresponding first and second sensor channels are in a normal operating state, assign a zero value to the corresponding row and column intersection position in the mask matrix. If it is determined that at least one of the corresponding first and / or second sensor channels is in a data missing, communication interrupted, or hardware damaged state (i.e., single channel missing or both channels missing), assign a negative infinity value to the corresponding row and column intersection position in the mask matrix.

[0084] For example, in one embodiment, the assignment formula for the corresponding element in the mask matrix can be expressed as:

[0085] ;

[0086] in, The corresponding element in the mask matrix; , These are the status identifier values ​​for each sensor channel.

[0087] In specific implementation, the input independent channel feature vectors are mapped and transformed through the corresponding query weight matrix, key weight matrix, and value weight matrix to obtain the corresponding query matrix, key matrix, and value matrix, respectively. The dot product similarity between the query matrix and the transpose of the key matrix is ​​calculated to obtain the original attention matrix between channels. The original attention matrix, the physical constraint matrix constructed from the physical constraint bias, and the mask matrix constructed from the mask bias are summed and superimposed element by element to obtain the superimposed matrix. The superimposed matrix is ​​scaled, and the dynamically adjusted attention distribution weights are calculated through a normalized activation function. The value matrix is ​​then weighted and summed using the attention distribution weights to output the feature matrix.

[0088] Specifically, the independent channel feature matrices output from the preceding steps are read. In the self-attention layer of the inverse Transformer architecture, three learnable linear mapping weight matrices are pre-defined: the query weight matrix, the key weight matrix, and the value weight matrix. The independent channel feature matrices are right-multiplied linearly with these three weight matrices respectively to obtain the query matrix, key matrix, and value matrix. The generated key matrix is ​​transposed to obtain the transpose matrix. The query matrix and the transpose matrix are multiplied using standard matrix multiplication (dot product). After multiplication, a two-dimensional square matrix is ​​output as the original attention matrix. The currently calculated original attention matrix, the physical constraint matrix constructed in the preceding steps, and the mask matrix are retrieved. The row and column dimensions of the three matrices are unified, and tensor linear superposition is enabled within the attention layer. The values ​​of the three matrices at the same coordinate positions are directly added to obtain the superimposed matrix. To prevent the dot product value from being too large and causing gradient vanishing, each element in the superimposed mixture matrix is ​​divided by a scaling factor to obtain the scaled fraction. The scaled matrix is ​​input into the normalization activation function, and exponential normalization is calculated along the row dimension of the matrix. After the activation function is calculated, the final dynamic attention distribution matrix is ​​output. The normalized dynamic attention distribution matrix is ​​then multiplied by the initially generated value matrix. This multiplication is essentially implemented by using attention probability values ​​as weights to dynamically weight and sum the value vectors of all channels, ultimately outputting the feature matrix.

[0089] For example, in one embodiment, the query matrix can be represented as:

[0090] ;

[0091] in, For query matrix; The feature matrix is ​​for each independent channel; To query the weight matrix.

[0092] The key matrix can be represented as:

[0093] ;

[0094] in, The key matrix; The feature matrix is ​​for each independent channel; This is the key weight matrix.

[0095] The value matrix can be represented as:

[0096] ;

[0097] in, It is a value matrix; The feature matrix is ​​for each independent channel; The value weight matrix.

[0098] The characteristic matrix can be represented as:

[0099] ;

[0100] in, The characteristic matrix; For query matrix; The key matrix; It is a value matrix; The physical constraint matrix; It is a mask matrix; The preset high-dimensional feature mapping dimension.

[0101] S104. Aggregate the feature matrix, extract the comprehensive shared representation of the device, and output health prediction results for multiple components in the device in parallel based on the comprehensive shared representation.

[0102] Specifically, the integrated shared representation refers to the high-order semantic vector representing the overall global operating and health status of the entire device, obtained by aggregating the feature matrices of various sensor channels through aggregation operators (such as global pooling, multilayer perceptron reassembly, or feature flattening). The health prediction result refers to the assessment conclusions of the future health status of multiple different key core components within the device, simultaneously output through multi-parallel prediction output heads based on the aforementioned unified integrated shared representation. The health prediction results can include the following parts: the first part is the component-level health score, which is represented by a continuous scalar value within a preset range (such as 0 to 1) to linearly quantify the current performance degradation of the component under monitoring; the second part is the multi-level degradation status label, which is represented by the preset status classification results to qualitatively indicate whether the component under monitoring is in a normal state, a sub-healthy state, an early degradation state, or a severe failure state; the third part is the multi-mode failure probability distribution, which is represented by the predicted probability vector of the occurrence of various typical mechanical / electrical failure modes preset for the component under monitoring; and the fourth part is the remaining life (RUL) prediction parameter, which is represented by the predicted remaining time or remaining number of cycles that the component under monitoring can continue to operate safely before triggering the shutdown threshold.

[0103] In specific implementation, the feature matrix is ​​pooled, and feature condensation and dimensionality reduction are performed in the channel dimension to extract a comprehensive shared representation vector. The comprehensive shared representation vector represents the current operating status of the entire device. Multiple parallel health prediction branches are constructed, each corresponding to a different monitored component in the device. The comprehensive shared representation vector is synchronously distributed and input into each health prediction branch. Within each health prediction branch, the comprehensive shared representation vector is sequentially subjected to feature mapping and transformation processing, and the health prediction results of each corresponding monitored component are output in parallel.

[0104] Optionally, when training the health prediction branches, the internal parameters of each health prediction branch are updated iteratively using a joint loss function that includes multi-task constraints. The joint loss function is composed of a weighted combination of the classification cross-entropy loss of each health prediction branch and the physical consistency loss used to penalize prediction results that violate common sense physics.

[0105] Specifically, feature condensation is performed on the channel dimension of the feature matrix (i.e., the row dimension C of the matrix). Global average pooling or global max pooling operators are used to aggregate and concatenate the feature values ​​of all channels along the row direction. After the pooling operation, the spatial dimension constraint of the number of channels C is eliminated, and the two-dimensional feature matrix is ​​collapsed and reduced to a one-dimensional dense tensor, thus extracting the comprehensive shared representation vector. At the back end of the deep learning model, N spatially parallel health prediction branches (classification output heads) are constructed, where the value of N corresponds one-to-one with the total number of components to be monitored in the current device (such as core components like main bearings, impellers, and motors). The previously extracted global comprehensive shared representation vector is used as a common underlying feature source, and through tensor replication and broadcasting mechanisms, it is synchronously distributed and input to these N independent health prediction branches, achieving decoupling of downstream multi-tasks while maintaining the sharing of underlying mechanism features. Within each independent health prediction branch, the hardware computation flow sequentially executes the following operators: The input comprehensive shared representation vector is multiplied by a learnable weight matrix specific to that branch and used to identify the state of a particular component; a bias vector is then added, and a linear transformation is performed to calculate the component health status score vector. This component health status score vector is then input into a preset feature mapping and activation module for transformation processing. As an optional implementation, if the health prediction task is a discrete state multi-class or multi-label recognition task, a probability transformation is performed using a normalized activation function: if the state classification is a mutually exclusive multi-class task (such as mutually exclusive health level grading), it is input into the Softmax function for row-level exponential normalization; if the state monitoring is a multi-label superimposed task or a binary classification task (such as composite abnormal state recognition), it is input into the Sigmoid function for independent activation. If the health prediction task is a quantitative continuous numerical regression task (such as component degradation scoring or remaining life prediction), the feature mapping and activation module performs a nonlinear or linear regression transformation to directly map the target regression scalar value. After the transformation or activation calculation is completed, the N health prediction branches simultaneously and independently output the specific health prediction results of each corresponding monitored component (including but not limited to: component health status probability distribution, multi-level degradation status label, health score and remaining lifetime parameters).

[0106] When iteratively training the constructed health prediction branch and the preceding network, the internal learnable parameters are updated collaboratively through the following steps: The probability results output in parallel by each health prediction branch are extracted and compared with the actual component labels to calculate the independent classification cross-entropy loss for each branch. A physical consistency verification module is constructed, synchronously inputting the health prediction results output by each branch into a preset industrial mechanism constraint equation. If the prediction results of different components violate mechanical engineering or aerodynamic physics common sense, a penalty mechanism is triggered, and the corresponding physical consistency loss is calculated and output. Preset task weighting coefficients are introduced (e.g., a weighting coefficient of 0.4 for classification cross-entropy loss and 0.6 for physical consistency loss), and the classification loss of each branch and the overall physical consistency loss are linearly weighted and summed to construct a joint loss function containing multi-task constraints. In the backpropagation calculation of the network, the global gradient is calculated based on this unified joint loss function, and the internal parameters of each health prediction branch and the preceding feature fusion network are iteratively updated collaboratively, ensuring that the final trained parameters can balance data-driven classification accuracy with real physical mechanism common sense.

[0107] For example, in one embodiment, the synthesized shared representation vector can be represented as:

[0108] ;

[0109] in, For the comprehensive shared representation vector; This is the characteristic matrix.

[0110] Health prediction results can be expressed as:

[0111] ;

[0112] in, This is the distribution vector for health prediction; For the first Learnable branch weight matrix for each health prediction branch; For the comprehensive shared representation vector; This is the bias vector.

[0113] The joint loss function can be expressed as:

[0114] ;

[0115] in, For the joint loss function; , Weighting coefficients for tasks; For classification cross-entropy loss; This represents a loss of physical consistency. The number of branches for predicting health; Index for the health prediction branch.

[0116] Optionally, after outputting health prediction results for multiple components in the device, the method further includes: collecting comprehensive shared representations of multiple associated devices in the current system to construct a device feature set; the device feature set represents the state of the multi-device cluster; calculating weights on the device feature set to output the dynamic fusion weights corresponding to each associated device in the system; using the dynamic fusion weights corresponding to each associated device, performing weighted summation and fusion on their respective comprehensive shared representations to generate a global comprehensive feature vector; the global comprehensive feature vector represents the overall health status of the multi-device cluster; and performing a system-level health assessment based on the global comprehensive feature vector to output global operation and maintenance decision results for the multi-device cluster.

[0117] Specifically, after the health prediction results of each independent device are output, the comprehensive shared representation vectors of M interconnected devices in the current system are collected in real time and in parallel through the system-level data bus or cluster network interface. The collected comprehensive shared representation vectors of all interconnected devices are concatenated and stacked or spliced, and uniformly stored in a feature matrix or dynamic buffer to construct a device feature set representing the current state of the multi-device cluster. The device feature set is input into a preset weight calculation module (such as a self-attention mechanism module, a correlation measurement network, or a multilayer perceptron) to calculate the interaction importance score of each interconnected device in the current multi-device cluster system. The importance scores are normalized using a normalized activation function, mapping and outputting the dynamic fusion weights corresponding to each interconnected device in the system. The dynamic fusion weights corresponding to each interconnected device are multiplied by their own comprehensive shared representation vector to obtain a weighted and scaled device feature vector. All weighted and scaled feature vectors of interconnected devices are element-wise concatenated and summed to form a one-dimensional dense global comprehensive feature vector. The generated global integrated feature vector is input into a pre-defined system-level health assessment backend network (such as a fully connected decision layer or a multilayer perceptron classifier) ​​for global feature recombination and nonlinear mapping transformation. After classification activation or multi-label output processing by the assessment backend network, global operation and maintenance decision results for multi-device clusters are output in parallel (including cluster system-level health score, collaborative maintenance suggestion scheme, load dynamic scheduling instruction or multi-device global early warning level).

[0118] For example, in one embodiment, the set of device features of the current state of a multi-device cluster can be represented as:

[0119] ;

[0120] in, This is a set of device characteristics representing the current state of a multi-device cluster. For the first The overall shared representation vector of each device.

[0121] The dynamic fusion weights can be expressed as:

[0122] ;

[0123] in, For the first The dynamic fusion weight of each associated device in the cluster system; No. The raw score of the interaction importance of each associated device in the current multi-device system.

[0124] The global comprehensive feature vector can be represented as:

[0125] ;

[0126] in, This is the global comprehensive feature vector; For the first The dynamic fusion weight of each associated device in the cluster system; For the first A comprehensive shared representation vector for each device; This refers to the number of devices.

[0127] The method provided in this embodiment, firstly, overturns the drawback of the traditional Transformer architecture, which only packages data by time steps and ignores cross-sensor spatial interaction. It adopts an inverse architecture, using the complete normalized time series of a single sensor channel as a feature unit (Token) for self-attention computation. Innovatively, during the self-attention matrix operation, it adds and superimposes a physical constraint bias matrix representing the mechanistic correlation between channels and a mask bias matrix representing the dynamic missing state of channels, thereby outputting a feature matrix that deeply encodes the spatial coupling relationship between different sensors. Secondly, it amplifies the attention weight of strongly correlated channels by assigning preset positive feature correlation constants at the intersection positions of the physical constraint matrices. By rigidly embedding operational mechanism knowledge such as mechanical topology and aerodynamics into the matrix calculation of deep learning networks, the black box drawbacks of pure data-driven approaches are effectively overcome. This forcibly suppresses spurious correlations and occasional overfitting caused by noise, achieving a dual improvement in the interpretability and accuracy of the diagnostic model. On the other hand, for frequent channel loss anomalies such as hardware damage or communication interruption, an efficient online dynamic fault tolerance and numerical suppression mechanism is provided by acquiring available states online and assigning negative infinity values ​​to the corresponding coordinates of the mask matrix. This mechanism can isolate and eliminate interference from defective channels in real time during the attention score calculation stage, preventing outliers from polluting global features and giving the entire system extremely strong structural noise resistance and hardware toughness.

[0128] Secondly, the standardized time-series data of each sensor channel are sequentially processed by one-dimensional convolution, activation, and pooling to extract time feature codes containing time-domain statistics and frequency-domain features. Channel codes and physical attribute codes are generated by embedding mapping according to the channel index and the type of monitored physical component. Finally, the three codes are summed, element-aligned, and fused to output independent channel feature vectors. By explicitly injecting prior multi-dimensional semantics such as spatial topology identifiers and core device structural attributes at the token level, the initial feature representation of each sensor channel has extremely strong spatiotemporal discernibility and physical correlation. This completely changes the limitation of traditional methods that rely on only a single waveform. It prevents semantic confusion caused by different dimensions when heterogeneous multi-source data is directly sent into the attention layer, thus laying a high-value, standardized spatial feature foundation for subsequent deep mining of cross-sensor spatial coupling.

[0129] Thirdly, based on the parallel output of single-device multi-component health prediction results, the system further elevates its dimensionality by collecting the comprehensive shared representation vectors of multiple related devices in the system at the current moment to construct a device feature set. The preset module calculates the dynamic fusion weights of each related device in the system, and these weights are used to weight and sum the comprehensive shared representations to generate a global comprehensive feature vector. Based on this vector, a system-level health assessment is performed, and the global operation and maintenance decision results for the multi-device cluster are output in parallel. This achieves a dimensional leap from single-point device isolation and early warning to multi-device cluster federated collaborative operation and maintenance. The dynamic fusion weights accurately capture the dynamic correlations and interaction importance between different devices within the cluster system, eliminating the local optimum bottlenecks and strategy conflicts caused by isolated single-device assessments in traditional methods. It can directly output cluster-level system-level health scores, collaborative maintenance suggestions, dynamic load scheduling instructions, or global early warning levels for multiple devices from a high-dimensional cluster perspective, providing globally optimal overall planning decisions. While maximizing feature sharing at the underlying level and reducing the massive computational overhead of redundant network construction, it solves the practical operation and maintenance challenges of strong coupling and collaborative evolution of multiple devices in large-scale rail transit or modern industrial scenarios.

[0130] Corresponding to the aforementioned embodiment of a multi-sensor, multi-component health prediction method, this application also provides an embodiment of a multi-sensor, multi-component health prediction device.

[0131] Figure 2 This is a schematic diagram of the multi-sensor, multi-component health prediction device provided in Embodiment 2 of this application. Please refer to... Figure 2 The apparatus provided in this embodiment includes a construction module and an output module;

[0132] The construction module is used to collect multi-source sensor data from different components in the device, perform time alignment and preprocessing on the multi-source sensor data, and construct a multi-channel data matrix with a unified spatial structure.

[0133] The construction module is also used to fuse time-series features, channel identification features, and physical component attribute features for each sensor channel in the multi-channel data matrix to construct a corresponding independent channel feature vector;

[0134] The output module is used to take the independent channel feature vectors of each sensor channel as input to the inverse Transformer architecture, and perform self-attention calculation with the sensor channel as the feature unit. In the self-attention calculation process, physical constraint biases that characterize the mechanism relationship between channels and mask biases that characterize the dynamic missing state of channels are introduced, and the output is a feature matrix that encodes the spatial coupling relationship between different sensors.

[0135] The output module is also used to aggregate the feature matrix, extract the comprehensive shared representation of the device, and output health prediction results for multiple components in the device in parallel based on the comprehensive shared representation.

[0136] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0137] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0138] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0139] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A multi-sensor, multi-component health prediction method, characterized in that, The method includes: Collect multi-source sensor data from different components in the device, perform time alignment and preprocessing on the multi-source sensor data, and construct a multi-channel data matrix with a unified spatial structure; For each sensor channel in the multi-channel data matrix, time-series features, channel identification features, and physical component attribute features are fused to construct a corresponding independent channel feature vector; The independent channel feature vectors of each sensor channel are used as input to the inverse Transformer architecture. Self-attention calculation is performed with the sensor channel as the feature unit. In the process of self-attention calculation, physical constraint biases that characterize the mechanism relationship between channels and mask biases that characterize the dynamic missing state of channels are introduced. The output is a feature matrix that encodes the spatial coupling relationship between different sensors. The feature matrix is ​​aggregated, a comprehensive shared representation of the device is extracted, and health prediction results for multiple components in the device are output in parallel based on the comprehensive shared representation.

2. The method according to claim 1, characterized in that, For each sensor channel in the multi-channel data matrix, time-series features, channel identifier features, and physical component attribute features are fused to construct a corresponding independent channel feature vector, including: The standardized time series data of each sensor channel are sequentially subjected to one-dimensional convolution, activation and pooling to extract time feature codes containing time domain statistical features and frequency domain features; An embedding mapping is performed based on the channel index of each sensor channel to generate a channel code that distinguishes the spatial location of different channels; Embedding mapping is performed based on the component type monitored by each sensor channel to generate physical attribute codes that characterize the specificity of different physical components; The time feature encoding, the channel encoding, and the physical attribute encoding are summed, element-aligned, and fused to output an independent channel feature vector.

3. The method according to claim 1, characterized in that, The independent channel feature vectors of each sensor channel are used as input to the inverse Transformer architecture. Self-attention calculation is performed using sensor channels as feature units. During the self-attention calculation process, physical constraint biases representing the mechanistic relationship between channels and mask biases representing the dynamic missing state of channels are introduced. The output is a feature matrix encoding the spatial coupling relationship between different sensors, including: Each input independent channel feature vector is mapped and transformed through the corresponding query weight matrix, key weight matrix, and value weight matrix to obtain the corresponding query matrix, key matrix, and value matrix, respectively. Calculate the dot product similarity between the query matrix and the transpose of the key matrix to obtain the original attention matrix between channels; The original attention matrix, the physical constraint matrix constructed from the physical constraint bias, and the mask matrix constructed from the mask bias are summed and superimposed element by element to obtain the superimposed matrix. The superimposed matrix is ​​scaled, and the dynamically adjusted attention distribution weights are calculated using a normalized activation function. The attention distribution weights are then used to perform a weighted summation of the value matrix to output the feature matrix.

4. The method according to claim 3, characterized in that, The physical constraint matrix, constructed from physical constraint biases, includes: Construct a square matrix whose number of rows and columns are equal to the total number of sensor channels as the physical constraint matrix; Based on the operating mechanism of the device, analyze whether there is a physical coupling relationship between any two sensor channels; If the corresponding first sensor channel and second sensor channel are related in physical structure or operating mechanism, a preset positive feature association constant is assigned to the intersection position of the first sensor channel and the second sensor channel in the physical constraint matrix according to the strength of the physical association between the first sensor channel and the second sensor channel; the positive feature association constant is used to amplify the attention weight between the first sensor channel and the second sensor channel. If there is no physical mechanism relationship between the corresponding first sensor channel and the second sensor channel, a zero value is assigned to the corresponding intersection position in the physical constraint matrix.

5. The method according to claim 3, characterized in that, The mask matrix constructed from the mask bias includes: Construct a square matrix whose number of rows and columns are equal to the total number of sensor channels as the mask matrix; Based on the current availability status of each sensor channel, obtain the status identifier value corresponding to each sensor channel; If both the corresponding first sensor channel and the second sensor channel are in normal condition, a zero value is assigned to the intersection position of the first sensor channel and the second sensor channel in the mask matrix; If at least one of the corresponding first sensor channel and / or second sensor channel is in a data missing state, a negative infinity value is assigned to the corresponding intersection position in the mask matrix.

6. The method according to claim 1, characterized in that, Aggregate the feature matrix, extract the comprehensive shared representation of the device, and output health prediction results for multiple components in the device in parallel based on the comprehensive shared representation, including: The feature matrix is ​​pooled to perform feature condensation and dimensionality reduction along the channel dimension, and a comprehensive shared representation vector is extracted; the comprehensive shared representation vector represents the current operating state of the entire device. Multiple parallel health prediction branches are constructed, each corresponding to a different monitored component in the device. The integrated shared representation vector is synchronously distributed and input to each health prediction branch. Within each health prediction branch, the comprehensive shared representation vector is sequentially subjected to feature mapping and transformation processing, and the health prediction results of each corresponding monitored component are output in parallel.

7. The method according to claim 6, characterized in that, The method includes: when training the health prediction branches, using a joint loss function containing multi-task constraints to collaboratively and iteratively update the internal parameters of each health prediction branch, wherein the joint loss function is composed of a weighted combination of the classification cross-entropy loss of each health prediction branch and the physical consistency loss used to penalize prediction results that violate physical common sense.

8. The method according to claim 1, characterized in that, The multi-source sensor data is time-aligned and preprocessed to construct a multi-channel data matrix with a unified spatial structure, including: The multi-source sensor data are arranged spatially according to a preset channel order to construct an initial multi-channel data matrix; For the initial multi-channel data matrix, the signal of the first channel is downsampled, the signal of the second channel is interpolated, and the data of all sensor channels are aligned to a unified time window for resampling; the sampling rate of the first channel is greater than that of the second channel. Calculate the statistical mean and standard deviation of each sensor channel within the time window after resampling. Subtract the corresponding statistical mean from the resampled data value of each channel and then divide by the corresponding standard deviation to output a standardized multi-channel data matrix. Based on the data integrity status of each sensor channel at the current moment, a corresponding mask vector is constructed, and the standardized multi-channel data matrix and the mask vector are output synchronously.

9. The method according to claim 1, characterized in that, After outputting health prediction results for multiple components in the device, the method further includes: Collect comprehensive shared representations of multiple associated devices within the current system to construct a device feature set; the device feature set represents the state of the multi-device cluster. The weights of the device feature set are calculated, and the dynamic fusion weights of each associated device in the system are output. By utilizing the dynamic fusion weights corresponding to each of the associated devices, the respective comprehensive shared representations are weighted and fused to generate a global comprehensive feature vector; the global comprehensive feature vector represents the overall health status of the multi-device cluster. Based on the global comprehensive feature vector, a system-level health assessment is performed, and a global operation and maintenance decision result is output for the multi-device cluster.

10. A multi-sensor, multi-component health prediction device, characterized in that, The device includes a construction module and an output module; The construction module is used to collect multi-source sensor data from different components in the device, perform time alignment and preprocessing on the multi-source sensor data, and construct a multi-channel data matrix with a unified spatial structure. The construction module is also used to fuse time-series features, channel identification features, and physical component attribute features for each sensor channel in the multi-channel data matrix to construct a corresponding independent channel feature vector; The output module is used to take the independent channel feature vectors of each sensor channel as input to the inverse Transformer architecture, and perform self-attention calculation with the sensor channel as the feature unit. In the self-attention calculation process, physical constraint biases that characterize the mechanism relationship between channels and mask biases that characterize the dynamic missing state of channels are introduced, and the output is a feature matrix that encodes the spatial coupling relationship between different sensors. The output module is also used to aggregate the feature matrix, extract the comprehensive shared representation of the device, and output health prediction results for multiple components in the device in parallel based on the comprehensive shared representation.

Images