A multi-mode fusion-based multi-working-condition life prediction method and system

Abstract

The application relates to the technical field of equipment life prediction, and relates to a multi-working-condition life prediction method based on multi-modal fusion, which comprises the following steps: obtaining equipment operation monitoring data, confirming a working condition setting sequence and a multi-channel sensor value sequence based on the equipment operation monitoring data, performing standardization processing on the multi-channel sensor value sequence to obtain a standardized sensor value sequence, constructing a working condition clustering center set based on the working condition setting sequence, performing working condition identification distribution on a working condition setting window by using the working condition clustering center set to obtain a working condition identification sequence, respectively performing local single-channel convolution and global cross-channel convolution on the working condition modulation features, splicing to obtain multi-scale time sequence features, and performing gated fusion on the attention hidden representation and the state space hidden representation to obtain a fused degradation hidden representation. The application can realize accurate residual life prediction under multiple working conditions.

Description

Technical Field

[0001] This invention relates to the field of equipment life prediction technology, and in particular to a multi-condition life prediction method based on multimodal fusion. Background Technology

[0002] In industrial equipment health management and predictive maintenance scenarios, remaining service life prediction is an important technical means to ensure safe equipment operation, optimize maintenance plans, and reduce operation and maintenance costs. As complex equipment such as aero engines, compressors, and rotating machinery operate under conditions of variable load, variable speed, and variable environmental parameters for extended periods, monitoring data has become characterized by multiple sensors, multi-source heterogeneity, and strong temporal coupling. Traditional life assessment methods oriented towards single stable operating conditions are no longer sufficient to meet the requirements for high-precision prediction.

[0003] Most existing lifetime prediction schemes are based directly on sensor time-series data for modeling, or simply stitch together operating condition information, lacking a dedicated assessment of the impact of changes in operating conditions. When equipment operating states frequently switch, sensor responses often simultaneously reflect operating condition disturbances and actual performance degradation, easily leading the model to misinterpret short-term operating condition fluctuations as lifetime decline trends. Furthermore, different sensor channels exhibit significant differences in their ability to characterize degradation; without effective channel selection and weight allocation mechanisms, redundant noise can be introduced, affecting prediction stability.

[0004] Furthermore, equipment degradation processes typically involve multi-scale features such as local anomalies, periodic fluctuations, and long-term slow decay. Existing methods still have shortcomings in temporal feature extraction and long-range dependency modeling, making it difficult to simultaneously represent local details and perceive global degradation trends. This results in low accuracy and insufficient generalization ability in predicting remaining service life under various operating conditions. Therefore, there is an urgent need for a technical solution that can integrate operating condition information with multi-channel sensor information, adapt to complex operating condition changes, and improve the accuracy of service life prediction. Summary of the Invention

[0005] This invention provides a multi-condition lifetime prediction method based on multimodal fusion and a computer-readable storage medium, the main purpose of which is to achieve accurate prediction of remaining lifetime under multiple conditions.

[0006] To achieve the above objectives, the present invention provides a multi-condition lifetime prediction method based on multi-modal fusion, comprising: Acquire equipment operation monitoring data, and based on the equipment operation monitoring data, determine the operating condition setting sequence and the multi-channel sensor value sequence; The multi-channel sensor numerical sequence is standardized to obtain a standardized sensor numerical sequence. The sensor window and the operating condition setting window are constructed using the standardized sensor numerical sequence and the operating condition setting sequence. A set of operating condition cluster centers is constructed based on the operating condition setting sequence. The operating condition cluster centers are used to assign operating condition identifiers to the operating condition setting window to obtain an operating condition identifier sequence. Operating condition statistical nodes are obtained based on the operating condition identifier sequence. The operating condition statistical nodes and the operating condition identifier sequence are used to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features. Local single-channel convolution and global cross-channel convolution are performed on the aforementioned operating condition modulation features respectively, and the features are concatenated to obtain multi-scale temporal features. Temporal pooling and channel-gated weighting are then applied to the multi-scale temporal features to obtain aggregated degradation features. The aggregated degradation features are input into the pre-constructed attention encoding branch and the pre-constructed state space encoding branch to obtain the attention hidden representation and the state space hidden representation. The attention hidden representation and the state space hidden representation are then gated and fused to obtain the fused degradation hidden representation. Based on the fused degradation hidden representation, the predicted value of the remaining service life of the device is obtained.

[0007] Optionally, the step of standardizing the multi-channel sensor numerical sequence to obtain a standardized sensor numerical sequence, and constructing a sensor window and a condition setting window using the standardized sensor numerical sequence and the operating condition setting sequence, includes: Multiple sensor channels and sensor value sets were identified based on multi-channel sensor numerical sequences; The multi-channel sensor numerical sequence is divided according to multiple sensor channels to obtain multiple channel numerical sequences; Perform the following operation on each channel of the multi-channel numerical sequence: Calculate the mean and standard deviation of the channel numerical sequences respectively to obtain the channel mean and channel standard deviation; By correlating the channel mean and channel standard deviation, channel standardized parameter nodes are obtained. By summarizing the channel standardized parameter nodes, a set of channel standardized parameter nodes is obtained, wherein each channel standardized parameter node corresponds one-to-one with a sensor channel. The sensor value set is standardized by using the channel standardized parameter node set, and then divided according to multiple sensor channels to obtain multiple standardized channel value sequences. The standardized sensor value sequence is obtained by sorting the multiple standardized channel value sequences according to the channel order of the multiple sensor channels in the multi-channel sensor value sequence.

[0008] Obtain the window length parameter, and based on the window length parameter, slide to extract continuous segments in the standardized sensor value sequence and the operating condition setting sequence to obtain the sensor window and the operating condition setting window.

[0009] Optionally, the step of constructing a set of operating condition cluster centers based on the operating condition setting sequence, and using the set of operating condition cluster centers to assign operating condition identifiers to the operating condition setting window to obtain an operating condition identifier sequence, includes: Multiple time points are obtained from the collection of equipment operation monitoring data, resulting in multiple collection time points; Obtain working condition setting vectors corresponding to multiple collection time points to obtain a training working condition vector set. Perform mean clustering on the training working condition vector set to obtain a working condition cluster center set, wherein the working condition cluster center set includes multiple working condition cluster centers. In the operating condition setting window, extract the window operating condition setting vectors sequentially according to the chronological order of the data acquisition time points, and perform the following operations on the extracted window operating condition setting vectors: Calculate the vector distance between the working condition setting vector and each working condition cluster center in the working condition cluster center set to obtain a vector distance set, which includes multiple vector distances; The working condition cluster center corresponding to the smallest vector distance in the vector distance set is identified to obtain the target working condition center. Based on the target working condition center, the working condition identifier is assigned to the collection time point corresponding to the working condition setting vector to obtain the time point working condition identifier. By summarizing the operating condition identifiers at the aforementioned time points, a sequence of operating condition identifiers is obtained.

[0010] Optionally, the step of obtaining operating condition statistical nodes based on the operating condition identifier sequence, and using the operating condition statistical nodes and the operating condition identifier sequence to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features includes: Obtain the latest time point and the latest time point sequence number from the equipment operation monitoring data; Input the time point working condition identifier corresponding to the latest time point into the pre-constructed working condition embedding mapping layer to obtain the working condition embedding vector. Input the working condition embedding vector into the pre-constructed multilayer sensing network and perform activation processing to obtain the working condition projection vector. The total number of times the operating condition identifier changes between adjacent time points in the statistical operating condition identifier sequence is obtained as the number of operating condition switching. The ratio of the number of operating condition switching to the latest time point sequence number is calculated to obtain the operating condition switching frequency. The number of times the working condition identifier corresponding to the latest time point has appeared in the working condition identifier sequence is counted to obtain the cumulative number of working conditions. The ratio of the cumulative number of working conditions to the latest time point sequence number is calculated to obtain the window working condition percentage. By associating the window condition ratio and condition switching frequency, a condition statistics node is obtained. The condition statistics node is then input into a pre-constructed learnable gating layer to obtain a channel modulation intensity vector. The channel modulation bias is obtained by multiplying the channel modulation intensity vector with the operating condition projection vector. The sensor vector at the latest time point in the sensor window is added to the channel modulation bias to obtain the modulation sensor vector at the latest time point. By summarizing the modulation sensor vectors, the modulation characteristics under operating conditions are obtained.

[0011] Optionally, the step of inputting the operating condition statistics node into a pre-constructed learnable gating layer to obtain the channel modulation intensity vector is calculated as follows:

[0012] in, This represents the channel modulation intensity vector. This indicates that the pre-built sigmoid activation function in the learnable gated layer is used. This indicates the pre-defined gate weight matrix in the learnable gate layer. This indicates the percentage of window conditions at the latest time point. This indicates the frequency of operating condition switching at the latest point in time. This represents the preset gate bias vector in the learnable gate layer. This indicates a vector concatenation operation.

[0013] Optionally, the step of performing local single-channel convolution and global cross-channel convolution on the modulated features of the operating conditions, and concatenating them to obtain multi-scale temporal features, includes: Obtain a set of relative scale coefficients, wherein the set of relative scale coefficients includes multiple relative scale coefficients, and generate multiple scale convolution kernels based on the set of relative scale coefficients and the window length parameter; For each of the multiple scale convolution kernels, perform the following operation: The scale convolution kernel is used to perform channel-by-channel independent depth convolution operation on the working condition modulation feature to obtain local single-channel convolution feature, and the scale convolution kernel is used to perform cross-channel standard convolution operation on the working condition modulation feature to obtain global cross-channel convolution feature. By summarizing local single-channel convolutional features and global cross-channel convolutional features and performing dimensional concatenation, multi-scale temporal features are obtained.

[0014] Optionally, the step of performing time-dimensional pooling and channel-gated weighting on the multi-scale temporal features to obtain aggregated degenerate features includes: Perform time-dimensional adaptive average pooling on multi-scale temporal features to obtain channel description vectors; The channel description vector is input into a pre-constructed two-layer fully connected gated network to obtain the channel importance weight vector; The channel importance weight vector is broadcast over time to obtain the channel weight time series matrix. The channel weight time series matrix is ​​then multiplied channel by channel by multi-scale time series features to obtain aggregated degradation features.

[0015] Optionally, the step of inputting the aggregated degenerate features into a pre-constructed attention encoding branch and a pre-constructed state space encoding branch to obtain an attention latent representation and a state space latent representation, and then performing gated fusion on the attention latent representation and the state space latent representation to obtain a fused degenerate latent representation, includes: A relative time stamp sequence is generated based on the window length parameter. The aggregated degradation feature and the relative time stamp sequence are input into the pre-constructed feature case embedding mapping layer in the attention encoding branch to obtain the encoder input embedding. The encoder input is then embedded into the pre-constructed multi-layer sparse self-attention encoder to obtain the attention hidden representation. The aggregated degenerate features are input into the pre-constructed selective state space model in the state space coding branch to obtain the state space latent representation; The attention latent representation and the state space latent representation are gated and fused to obtain the fused degenerate latent representation. The calculation formula for the gated fusion is as follows:

[0016]

[0017] in, Indicates the gating fusion coefficient. This represents the preset fusion layer weight matrix. This represents the attention hidden representation. This represents the implicit representation of the state space. This represents the preset fusion layer bias vector. This represents the fusion-degenerate hidden representation. This indicates an element-wise multiplication operation.

[0018] Optionally, obtaining the predicted remaining useful life of the device based on the fused degradation latent representation includes: Extract the terminal history fragment from the fused degenerate hidden representation, and concatenate the terminal history fragment with the preset prediction placeholder vector to obtain the decoder input sequence; The hidden state at the predicted location is obtained by performing cross-attention calculation on the decoder input sequence and the attention hidden representation using a pre-defined decoder. A linear projection operation is performed on the predicted hidden state of the location to obtain the predicted value of the remaining service life of the device.

[0019] To achieve the above objectives, the present invention also provides a multi-condition lifetime prediction system based on multimodal fusion, comprising: The data preprocessing module is used to acquire equipment operation monitoring data and, based on the equipment operation monitoring data, determine the operating condition setting sequence and the multi-channel sensor value sequence. The multi-channel sensor numerical sequence is standardized to obtain a standardized sensor numerical sequence. The sensor window and the operating condition setting window are constructed using the standardized sensor numerical sequence and the operating condition setting sequence. The operating condition modulation module is used to construct an operating condition cluster center set based on the operating condition setting sequence, use the operating condition cluster center set to allocate operating condition identifiers to the operating condition setting window to obtain an operating condition identifier sequence, obtain operating condition statistical nodes based on the operating condition identifier sequence, and use the operating condition statistical nodes and the operating condition identifier sequence to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features. The feature extraction module is used to perform local single-channel convolution and global cross-channel convolution on the modulated features of the working conditions respectively, and concatenate them to obtain multi-scale temporal features. The multi-scale temporal features are then subjected to temporal pooling and channel-gated weighting to obtain aggregated degradation features. The lifetime prediction module is used to input the aggregated degradation features into a pre-constructed attention encoding branch and a pre-constructed state space encoding branch to obtain attention latent representation and state space latent representation, perform gating fusion on the attention latent representation and state space latent representation to obtain fused degradation latent representation, and obtain the predicted value of the remaining lifetime of the device based on the fused degradation latent representation.

[0020] To address the above problems, the present invention also provides an electronic device, the electronic device comprising: Memory, storing at least one instruction; The processor executes the instructions stored in the memory to implement the multi-condition lifetime prediction method based on multi-modal fusion described above.

[0021] To address the aforementioned problems, the present invention also provides a computer-readable storage medium storing at least one instruction, which is executed by a processor in an electronic device to implement the multi-condition lifetime prediction method based on multimodal fusion described above.

[0022] To address the problems described in the background art, this invention acquires equipment operation monitoring data, identifies the operating condition setting sequence and multi-channel sensor numerical sequences based on the monitoring data, standardizes the multi-channel sensor numerical sequences to obtain standardized sensor numerical sequences, and constructs sensor windows and operating condition setting windows using the standardized sensor numerical sequences and operating condition setting sequences. It is evident that this invention eliminates differences in dimensions and numerical ranges between different sensor channels by calculating the mean and standard deviation for each channel, enabling the model to perform feature learning at a unified numerical scale. Based on the operating condition setting sequences, a set of operating condition cluster centers is constructed, and these cluster centers are used to analyze the operating conditions. The system assigns operating condition identifiers to a window to obtain an operating condition identifier sequence. Based on the operating condition identifier sequence, it obtains operating condition statistical nodes. Using the operating condition statistical nodes and the operating condition identifier sequence, it performs channel-level adaptive modulation on the sensor window to obtain operating condition modulation features. It can be seen that the present invention generates channel-level modulation signals based on the prefix statistics of the operating condition identifier sequence, enabling the model to obtain operating condition perception capabilities before extracting degradation features. It distinguishes the contributions of operating condition factors and degradation factors in sensor value changes. Local single-channel convolution and global cross-channel convolution are performed on the operating condition modulation features respectively, and the features are concatenated to obtain multi-scale temporal features. The multi-scale temporal features are then subjected to time-dimensional pooling and channel-gated weighting to obtain aggregated degradation features. As can be seen, this invention acquires degradation features at different temporal granularities and sensor association levels through multi-scale local and global convolutions, and adaptively adjusts the contribution of each channel using a channel gating mechanism, thereby improving the comprehensiveness and specificity of degradation feature representation. The aggregated degradation features are input into pre-constructed attention encoding branches and pre-constructed state space encoding branches to obtain attention latent representations and state space latent representations. These latent representations are then gated and fused to obtain a fused degradation latent representation. Based on this fused degradation latent representation, the predicted remaining lifespan of the device is obtained. Thus, this invention achieves accurate prediction of the remaining lifespan of the device by modeling in parallel with attention encoding branches and state space encoding branches, adaptively integrating global attention relationships and continuous state transmission information using a gating fusion mechanism, and guiding the global degradation trend to the prediction position using a decoder cross-attention structure. Therefore, this invention can achieve accurate prediction of remaining lifespan under multiple operating conditions. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a multi-condition lifetime prediction method based on multimodal fusion, provided in an embodiment of the present invention. Figure 2 A functional block diagram of a multi-condition lifetime prediction system based on multimodal fusion is provided in an embodiment of the present invention. Figure 3This is a schematic diagram of the structure of an electronic device that implements the multi-condition lifetime prediction method based on multi-modal fusion, according to an embodiment of the present invention.

[0024] Explanation of reference numerals in the attached figures: 10. Electronic device; 11. Processor; 12. Memory; 13. Bus.

[0025] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0026] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0027] This application provides a multi-condition lifetime prediction method based on multimodal fusion. The executing entity of this multimodal fusion-based multi-condition lifetime prediction method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the multimodal fusion-based multi-condition lifetime prediction method can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0028] Reference Figure 1 The diagram shown is a flowchart illustrating a multi-condition lifetime prediction method based on multimodal fusion according to an embodiment of the present invention. In this embodiment, the multi-condition lifetime prediction method based on multimodal fusion includes: S1. Obtain equipment operation monitoring data, and based on the equipment operation monitoring data, determine the operating condition setting sequence and the multi-channel sensor value sequence.

[0029] It should be explained that the equipment operation monitoring data refers to the monitoring records continuously collected and stored during equipment operation according to the acquisition time points. This data includes two types of information: operating condition setting parameters and multi-channel sensor measurement values. The equipment operation monitoring data is acquired as follows: during equipment operation, the operating condition parameters and sensor measurement values ​​are synchronously collected according to a fixed sampling period using a data acquisition device pre-deployed on the equipment. After collection, the data is sorted and stored according to the equipment number and acquisition time. The operating condition setting sequence is a sequence of operating condition setting vectors corresponding to each acquisition time point, arranged in chronological order. These vectors describe the current operating status setting parameters of the equipment, including but not limited to equipment operating height, operating Mach number, and throttle lever angle. The multi-channel sensor value sequence is a two-dimensional data sequence of values ​​measured by each sensor channel at each acquisition time point, arranged in chronological and channel order. Each acquisition time point corresponds to measurement values ​​from multiple sensor channels, including but not limited to temperature sensor channels, pressure sensor channels, and speed sensor channels.

[0030] It is understood that the equipment described in the embodiments of the present invention can be an aircraft engine, or a precision testing equipment used by testing and inspection institutions such as an octane number tester and a cetane number tester. The equipment has problems of switching between multiple operating conditions and sensor value drift during operation. The operating condition setting parameters can be fuel standard number, speed setting value, and oil bath temperature, etc.

[0031] For example, an aero-engine maintenance engineer needs to predict the remaining service life of a turbofan engine. First, equipment operation monitoring data of the turbofan engine is acquired through a data acquisition device. This data includes a sequence of operating condition settings consisting of three parameters: engine altitude, Mach number, and throttle lever angle at each acquisition time point, as well as a multi-channel sensor value sequence consisting of 21 sensor channels, including total temperature sensor channel, total pressure sensor channel, physical fan speed sensor channel, corrected fan speed sensor channel, high-pressure compressor outlet static pressure sensor channel, and turbine outlet temperature sensor channel. Similarly, engineers in inspection agencies can use this method to predict the remaining service life of octane or cetane number analyzers under different fuel grade testing conditions.

[0032] S2. Standardize the multi-channel sensor numerical sequence to obtain a standardized sensor numerical sequence. Use the standardized sensor numerical sequence and the operating condition setting sequence to construct a sensor window and an operating condition setting window.

[0033] It should be understood that the standardization process is an operation that performs a unified numerical scale transformation on the values ​​of each sensor channel in the multi-channel sensor numerical sequence, aiming to eliminate differences in dimensions and numerical ranges between different sensor channels. The standardized sensor numerical sequence is a numerical sequence obtained after standardization of the multi-channel sensor numerical sequence, with each channel value at the same numerical scale. The sensor window is a continuous segment obtained by sliding and cropping the standardized sensor numerical sequence according to a fixed window length, used as the input sample for the model. The operating condition setting window is a continuous segment within the same time range corresponding to the sensor window, cropped from the operating condition setting sequence. Because there are significant differences in the physical dimensions and numerical ranges of different sensor channels, it is necessary to standardize the values ​​of each channel to eliminate the influence of dimensions. Therefore, the standardization process of the multi-channel sensor numerical sequence to obtain a standardized sensor numerical sequence, and the construction of the sensor window and operating condition setting window using the standardized sensor numerical sequence and the operating condition setting sequence, includes: Multiple sensor channels and sensor value sets were identified based on multi-channel sensor numerical sequences; The multi-channel sensor numerical sequence is divided according to multiple sensor channels to obtain multiple channel numerical sequences; Perform the following operation on each channel of the multi-channel numerical sequence: Calculate the mean and standard deviation of the channel numerical sequences respectively to obtain the channel mean and channel standard deviation; By correlating the channel mean and channel standard deviation, channel standardized parameter nodes are obtained. By summarizing the channel standardized parameter nodes, a set of channel standardized parameter nodes is obtained, wherein each channel standardized parameter node corresponds one-to-one with a sensor channel. The sensor value set is standardized by using the channel standardized parameter node set, and then divided according to multiple sensor channels to obtain multiple standardized channel value sequences. The standardized sensor value sequence is obtained by sorting the multiple standardized channel value sequences according to the channel order of the multiple sensor channels in the multi-channel sensor value sequence.

[0034] Obtain the window length parameter, and based on the window length parameter, slide to extract continuous segments in the standardized sensor value sequence and the operating condition setting sequence to obtain the sensor window and the operating condition setting window.

[0035] Understandably, the sensor channel is the identifier of each sensor in the multi-channel sensor value sequence, used to distinguish the measurement data of different sensors. The sensor value set is the collection of values ​​from all acquisition time points and all sensor channels in the multi-channel sensor value sequence. The channel value sequence is a one-dimensional value sequence formed by arranging the measurement values ​​of a certain sensor channel in chronological order across all acquisition time points. The channel mean is the arithmetic mean of all values ​​in the channel value sequence. The channel standard deviation is a measure of the dispersion of all values ​​in the channel value sequence from the channel mean. The channel standardization parameter node is a data node composed of the channel mean and the channel standard deviation, used to standardize the values ​​of that sensor channel. The specific calculation method for the standardization transformation is as follows: subtract the channel mean of the corresponding sensor channel from each sensor value in the sensor value set, and then divide by the channel standard deviation of that sensor channel, thereby converting each channel value into a standardized value with a mean of zero and a standard deviation of one. The standardized channel value sequence is a sequence formed by arranging the standardized values ​​of a certain sensor channel in chronological order.

[0036] Furthermore, the window length parameter is the time length of the sliding window, used to determine how many consecutive acquisition time points of data are included in each input sample. The window length parameter can be set according to the device degradation cycle and model input requirements. The sliding interception is an operation that sequentially intercepts continuous segments starting from the beginning position of the standardized sensor value sequence and the operating condition setting sequence with a fixed window length parameter as the step size. The shape of the sensor window is a two-dimensional matrix of the window length parameter multiplied by the number of sensor channels. The shape of the operating condition setting window is a two-dimensional matrix of the window length parameter multiplied by the dimension of the operating condition setting vector.

[0037] For example, suppose the multi-channel sensor numerical sequence of a turbofan engine contains 21 sensor channels and a total of 192 acquisition time points. Taking the first sensor channel (total temperature sensor channel) as an example, this channel's numerical sequence contains 192 measurements. The calculated channel mean is 518.67, and the channel standard deviation is 0.85. The channel mean 518.67 and the channel standard deviation 0.85 are correlated to obtain the channel's standardized parameter node. Standardization is performed on each measurement belonging to the total temperature sensor channel in the sensor numerical set. For example, if the measurement value at a certain time point is 519.52, the standardized measurement value is (519.52-518.67) / 0.85=1.0. After performing the above operation on all 21 sensor channels, the standardized sensor numerical sequence is obtained by arranging them in channel order. The window length parameter is set to 30. Continuous segments are slidably extracted from the standardized sensor value sequence and the operating condition setting sequence to obtain the sensor window and the operating condition setting window. The sensor window has a shape of 30 x 21, and the operating condition setting window has a shape of 30 x 3. This embodiment of the invention performs standardization by calculating the mean and standard deviation for each channel, eliminating the differences in dimensions and numerical ranges between different sensor channels, enabling the model to perform feature learning at a uniform numerical scale.

[0038] S3. Construct a cluster center set of operating conditions based on the operating condition setting sequence, use the cluster center set of operating conditions to assign operating condition identifiers to the operating condition setting window to obtain an operating condition identifier sequence, obtain operating condition statistical nodes based on the operating condition identifier sequence, and use the operating condition statistical nodes and the operating condition identifier sequence to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features.

[0039] It should be explained that the operating condition cluster center set is a collection of multiple cluster centers obtained after clustering analysis of the operating condition setting vectors in the operating condition setting sequence. Each operating condition cluster center represents a typical operating condition mode. The operating condition identifier sequence is a sequence formed by arranging discrete operating condition identifiers in chronological order after assigning the operating condition setting vector at each acquisition time point in the operating condition setting window to the nearest operating condition cluster center. The operating condition statistical node is a data node of operating condition change characteristics obtained based on the statistical analysis of the operating condition identifier sequence. The channel-level adaptive modulation is an operation that uses the operating condition statistical node to adaptively adjust the value of each sensor channel in the sensor window, aiming to distinguish whether the sensor value change is caused by equipment degradation or by changes in operating conditions. The operating condition modulation feature is the feature data obtained after channel-level adaptive modulation of the sensor window, which already contains the operating condition adaptive adjustment information.

[0040] Understandably, equipment may experience various operating conditions during operation, and these different conditions can cause fluctuations in sensor values. These fluctuations, caused by changes in operating conditions, may interfere with the judgment of the true degradation trend of the equipment. Therefore, it is necessary to first identify the type of operating condition and then perform adaptive modulation of the sensor features based on the operating condition setting sequence. The process involves constructing a set of operating condition cluster centers based on the operating condition setting sequence, and then using this set to assign operating condition identifiers to the operating condition setting window, resulting in a sequence of operating condition identifiers, including: Multiple time points are obtained from the collection of equipment operation monitoring data, resulting in multiple collection time points; Obtain working condition setting vectors corresponding to multiple collection time points to obtain a training working condition vector set. Perform mean clustering on the training working condition vector set to obtain a working condition cluster center set, wherein the working condition cluster center set includes multiple working condition cluster centers. In the operating condition setting window, extract the window operating condition setting vectors sequentially according to the chronological order of the data acquisition time points, and perform the following operations on the extracted window operating condition setting vectors: Calculate the vector distance between the working condition setting vector and each working condition cluster center in the working condition cluster center set to obtain a vector distance set, which includes multiple vector distances; The working condition cluster center corresponding to the smallest vector distance in the vector distance set is identified to obtain the target working condition center. Based on the target working condition center, the working condition identifier is assigned to the collection time point corresponding to the working condition setting vector to obtain the time point working condition identifier. By summarizing the operating condition identifiers at the aforementioned time points, a sequence of operating condition identifiers is obtained.

[0041] It should be understood that the training operating condition vector set is a collection of operating condition setting vectors corresponding to all collection time points in the equipment operation monitoring data, used as training data for mean clustering. Mean clustering is an unsupervised clustering method that divides the training operating condition vector set into multiple clusters through iterative calculation. The center of each cluster is a cluster center. Optionally, the K-means clustering algorithm can be used to achieve this. The number of clusters in mean clustering is determined as follows: if the number of equipment operating conditions is known, the number of clusters is directly set to the known number of operating conditions. If the number of equipment operating conditions is unknown, the number of clusters can be determined by evaluating the training operating condition vector set using the elbow method or the silhouette coefficient method. The elbow method involves gradually increasing the number of clusters and calculating the total sum of squared errors under each cluster number, selecting the cluster number corresponding to the point where the rate of decrease in the total sum of squared errors slows significantly. The silhouette coefficient method calculates the average distance between each sample and other samples within its cluster and the average distance to the nearest neighbor sample under each cluster number, comparing the two values, and selecting the cluster number corresponding to the largest silhouette coefficient. The vector distance is the Euclidean distance between the operating condition setting vector and the operating condition cluster center. The target operating condition center is the operating condition cluster center that is closest to the current window operating condition setting vector, representing the operating condition type at the current acquisition time point. The operating condition identifier assignment is the operation of assigning the number of the target operating condition center to the current acquisition time point. The time point operating condition identifier is a discrete operating condition number assigned to a certain acquisition time point, used to identify the operating condition type to which that time point belongs.

[0042] Furthermore, to quantify the impact of changes in operating conditions on sensor characteristics, it is necessary to statistically analyze the frequency and switching patterns of operating conditions based on the operating condition identifier sequence. Therefore, the step of obtaining operating condition statistical nodes based on the operating condition identifier sequence, and using the operating condition statistical nodes and the operating condition identifier sequence to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features includes: Obtain the latest time point and the latest time point sequence number from the equipment operation monitoring data; Input the time point working condition identifier corresponding to the latest time point into the pre-constructed working condition embedding mapping layer to obtain the working condition embedding vector. Input the working condition embedding vector into the pre-constructed multilayer sensing network and perform activation processing to obtain the working condition projection vector. The total number of times the operating condition identifier changes between adjacent time points in the statistical operating condition identifier sequence is obtained as the number of operating condition switching. The ratio of the number of operating condition switching to the latest time point sequence number is calculated to obtain the operating condition switching frequency. The number of times the working condition identifier corresponding to the latest time point has appeared in the working condition identifier sequence is counted to obtain the cumulative number of working conditions. The ratio of the cumulative number of working conditions to the latest time point sequence number is calculated to obtain the window working condition percentage. By associating the window condition ratio and condition switching frequency, a condition statistics node is obtained. The condition statistics node is then input into a pre-constructed learnable gating layer to obtain a channel modulation intensity vector. The channel modulation bias is obtained by multiplying the channel modulation intensity vector with the operating condition projection vector. The sensor vector at the latest time point in the sensor window is added to the channel modulation bias to obtain the modulation sensor vector at the latest time point. By summarizing the modulation sensor vectors, the modulation characteristics under operating conditions are obtained.

[0043] Specifically, the latest time point is the latest acquisition time point in the time sequence within the sensor window. The latest time point sequence number is the position number of the latest time point in the operating condition identifier sequence. The operating condition embedding mapping layer is a learnable lookup table structure that converts discrete operating condition identifiers into low-dimensional continuous vector representations. During model training, the operating condition embedding mapping layer optimizes its parameters using a backpropagation algorithm. The operating condition embedding vector is a low-dimensional continuous vector obtained after the operating condition identifier at a given time point is transformed by the operating condition embedding mapping layer, used to represent the feature information of the current operating condition type. The multilayer sensing network is a feedforward neural network structure composed of multiple fully connected layers and activation functions, used to map the operating condition embedding vector to a dimensional space consistent with the number of sensor channels. The activation processing is an operation that uses the activation function to perform a nonlinear transformation on the output of the fully connected layer; optionally, the ReLU activation function can be used to achieve this purpose. The operating condition projection vector is a vector obtained after the operating condition embedding vector has passed through the multilayer sensing network and activation processing, with a dimension consistent with the number of sensor channels, used to generate the operating condition bias for each sensor channel.

[0044] Understandably, the confirmation method for the number of operating condition switching is as follows: In the operating condition identifier sequence, the operating condition identifiers of adjacent collection time points are compared sequentially. If two adjacent time point operating condition identifiers are different, it is counted once. The total number of times the operating condition identifiers of all adjacent time points change is counted, and this total number is the number of operating condition switching. The operating condition switching frequency is the ratio of the number of operating condition switching divided by the latest time point sequence number, reflecting the frequency of changes in the equipment's operating condition. The cumulative number of operating conditions is the total number of times the operating condition identifier corresponding to the latest time point appears in the operating condition identifier sequence. The window operating condition proportion is the ratio of the cumulative number of operating conditions divided by the latest time point sequence number, reflecting the proportion of a certain operating condition in the current window's history. The operating condition statistics node is a two-dimensional statistical data node composed of the window operating condition proportion and the operating condition switching frequency, used to describe the historical distribution characteristics of operating conditions within the current window.

[0045] It should be explained that the learnable gating layer is a gated network structure that uses the sigmoid activation function to map the input to a range of 0 to 1, used to generate the modulation intensity of each sensor channel. The gating layer weight matrix and gating layer bias vector in the learnable gating layer are obtained as follows: the output dimension of the gating layer weight matrix is ​​determined based on the number of sensor channels, and the input dimension is determined based on the dimension of the working condition statistics nodes. Then, the gating layer weight matrix is ​​randomly initialized using a normal distribution with a mean of 0 and a standard deviation of 0.01, and the gating layer bias vector is initialized to a zero vector. Iterative optimization is performed during the end-to-end training of the model using the backpropagation algorithm. The channel modulation intensity vector is obtained by inputting the working condition statistics nodes into the pre-constructed learnable gating layer, and the calculation formula is as follows:

[0046] in, This represents the channel modulation intensity vector. This indicates that the pre-built sigmoid activation function in the learnable gated layer is used. This indicates the pre-defined gate weight matrix in the learnable gate layer. This indicates the percentage of window conditions at the latest time point. This indicates the frequency of operating condition switching at the latest point in time. This represents the preset gate bias vector in the learnable gate layer. This indicates a vector concatenation operation.

[0047] It should be understood that the channel modulation intensity vector is a vector output by the learnable gating layer, with its dimension matching the number of sensor channels. Each element in the vector ranges from 0 to 1, and is used to control the intensity of the modulation received by the corresponding sensor channel under the operating conditions. The channel modulation bias is a vector obtained by element-wise multiplication of the channel modulation intensity vector and the operating condition projection vector, representing the amount of bias that needs to be superimposed on each sensor channel under the current operating conditions. The sensor vector is a one-dimensional vector composed of the standardized values ​​of all sensor channels at a certain time point in the sensor window, arranged in channel order, with a vector dimension equal to the number of sensor channels. The modulation sensor vector is a vector obtained by adding the sensor vector at a certain time point in the sensor window to the channel modulation bias. The operating condition modulation feature is feature data obtained by summarizing all modulation sensor vectors in chronological order after performing the above modulation operations at all time points in the sensor window; its shape is the same as the sensor window.

[0048] For example, suppose the operating condition setting sequence of a turbofan engine contains 6 operating conditions. K-means clustering is used to perform mean clustering on the training operating condition vector set, with the number of clusters set to 6, resulting in a set of operating condition cluster centers containing 6 operating condition cluster centers. The operating condition setting vector for a specific acquisition time point within the operating condition setting window is extracted as (1.0, 0.0042, 100.0). The Euclidean distance between this window operating condition setting vector and the 6 operating condition cluster centers is calculated, resulting in a vector distance set of (0.12, 2.35, 1.87, 3.01, 2.56, 1.42). The smallest vector distance is 0.12, corresponding to the first operating condition cluster center; therefore, the operating condition identifier for this acquisition time point is assigned a value of 1. The above operation is performed on all 30 acquisition time points within the window to obtain the operating condition identifier sequence. Assuming the latest time point number is 30, and the total number of times the operating condition identifier changes between adjacent time points in the operating condition identifier sequence is 4, then the operating condition switching frequency is 4 / 30 = 0.133. Assuming the operating condition identifier corresponding to the latest time point is 1, and the number of times operating condition identifier 1 has appeared in the operating condition identifier sequence is 12, then the window operating condition proportion is 12 / 30 = 0.4. Associating the window operating condition proportion of 0.4 with the operating condition switching frequency of 0.133 yields the operating condition statistics node (0.4, 0.133). This operating condition statistics node is input into a learnable gating layer, and after processing with a sigmoid activation function, a 21-dimensional channel modulation intensity vector is obtained. The latest time point's operating condition identifier 1 is input into the operating condition embedding mapping layer to obtain a 32-dimensional operating condition embedding vector. After processing with a multilayer perceptron and ReLU activation, a 21-dimensional operating condition projection vector is obtained. The channel modulation bias is obtained by element-wise multiplying the channel modulation intensity vector with the operating condition projection vector. The modulation sensor vector is obtained by adding the sensor vector at the latest time point in the sensor window to the channel modulation bias. After performing the above operations on all time points, the operating condition modulation features are summarized. This embodiment of the invention generates channel-level modulation signals based on prefix statistics of the operating condition identifier sequence, enabling the model to acquire operating condition perception capabilities before extracting degradation features, and distinguishing the respective contributions of operating condition factors and degradation factors in sensor numerical changes.

[0049] S4. Perform local single-channel convolution and global cross-channel convolution on the modulated features of the operating conditions respectively, and concatenate them to obtain multi-scale temporal features. Perform time-dimensional pooling and channel-gated weighting on the multi-scale temporal features to obtain aggregated degradation features.

[0050] It should be understood that the local single-channel convolution is a depthwise convolution operation performed independently on each sensor channel in the operational condition modulation features. That is, the convolution kernel slides only along the time dimension of a single sensor channel, without mixing data from other sensor channels, and is used to extract the degradation fluctuation features of each sensor channel at different time scales. The global cross-channel convolution is a standard convolution operation performed on the operational condition modulation features. That is, the convolution kernel calculates simultaneously along both the time and channel dimensions, integrating information from multiple sensor channels, and is used to extract system-level degradation patterns formed by common changes among multiple sensor channels. The multi-scale temporal features are feature data obtained by concatenating multiple local single-channel convolution features and global cross-channel convolution features along the channel dimension, containing degradation information at different time granularities and different sensor association levels. The aggregated degradation features are feature data obtained after evaluating and weighting the multi-scale temporal features based on channel importance. Channels that contribute significantly to degradation prediction are enhanced, while channels that contribute less are suppressed.

[0051] Understandably, equipment degradation processes include both short-term local abnormal fluctuations and long-term slow wear trends, requiring the extraction of degradation features across different time ranges using convolutional kernels of different scales. Therefore, the process of performing local single-channel convolution and global cross-channel convolution on the modulated features of the operating conditions, and then concatenating them to obtain multi-scale temporal features, includes: Obtain a set of relative scale coefficients, wherein the set of relative scale coefficients includes multiple relative scale coefficients, and generate multiple scale convolution kernels based on the set of relative scale coefficients and the window length parameter; For each of the multiple scale convolution kernels, perform the following operation: The scale convolution kernel is used to perform channel-by-channel independent depth convolution operation on the working condition modulation feature to obtain local single-channel convolution feature, and the scale convolution kernel is used to perform cross-channel standard convolution operation on the working condition modulation feature to obtain global cross-channel convolution feature. By summarizing local single-channel convolutional features and global cross-channel convolutional features and performing dimensional concatenation, multi-scale temporal features are obtained.

[0052] It should be explained that the relative scale coefficient set is a collection of scaling factors used to adaptively generate convolutional kernels of different sizes based on the window length parameter. The relative scale coefficient set is set by selecting three different scaling factors based on the time scale requirements of device degradation characteristics, corresponding to fine-scale, medium-scale, and coarse-scale time coverage, respectively. The size of the scale convolutional kernel is determined by both the relative scale coefficient and the window length parameter. Specifically, the calculation method is as follows: multiply the relative scale coefficient by the window length parameter, round the product down, multiply by two, and then add one to ensure that the convolutional kernel size is odd. This facilitates symmetrical padding to maintain the same time length of the output sequence. Symmetrical padding involves adding the same number of zero values ​​at the beginning and end of the input sequence, ensuring that the time steps of the output sequence after convolution are consistent with the input sequence. The depthwise convolution operation is a convolution operation where the number of groups equals the number of sensor channels, and each sensor channel uses an independent convolutional kernel for convolution calculation. The standard convolution operation is a conventional convolution operation where the convolutional kernel simultaneously covers all sensor channels. The local single-channel convolutional features are obtained by performing a depthwise convolution operation on the operating condition modulation features, preserving the temporal degradation pattern of each sensor channel. The global cross-channel convolutional features are obtained by performing a standard convolution operation on the operating condition modulation features, including the correlation degradation patterns between multiple sensor channels. The dimension concatenation is an operation that joins the local single-channel convolutional features and global cross-channel convolutional features of all scales along the channel dimension.

[0053] Furthermore, different sensor channels contribute differently to the prediction of the remaining useful life of the device. Some sensor channels are highly correlated with degradation, while others may approximate noise. Therefore, adaptive adjustment of channel contributions is required. The process of performing time-dimensional pooling and channel-gated weighting on the multi-scale time-series features to obtain aggregated degradation features includes: Perform time-dimensional adaptive average pooling on multi-scale temporal features to obtain channel description vectors; The channel description vector is input into a pre-constructed two-layer fully connected gated network to obtain the channel importance weight vector; The channel importance weight vector is broadcast over time to obtain the channel weight time series matrix. The channel weight time series matrix is ​​then multiplied channel by channel by multi-scale time series features to obtain aggregated degradation features.

[0054] It should be understood that the time-dimensional adaptive average pooling operation is an operation that calculates the average value of multi-scale temporal features over time, compressing the features of each channel within the entire time window into a scalar value, thereby obtaining a global statistical summary for each channel. The channel description vector is the vector output by the time-dimensional adaptive average pooling operation; the vector dimension equals the number of channels in the multi-scale temporal features, and each element represents the average feature intensity of the corresponding channel within the entire time window. The two-layer fully connected gated network is a lightweight network structure composed of two fully connected layers and activation functions. The first fully connected layer compresses the channel dimension to the original dimension divided by the dimensionality reduction ratio. After passing through the ReLU activation function, the second fully connected layer restores the dimension to the original channel dimension, and finally, after passing through the sigmoid activation function, outputs weight values ​​between zero and one. The dimensionality reduction ratio is a compression coefficient used to control the number of parameters in the gated network. When the number of channels in the multi-scale temporal features is large, a larger dimensionality reduction ratio is selected; otherwise, a smaller dimensionality reduction ratio is selected. The channel importance weight vector is the vector output by the two-layer fully connected gated network, and each element represents the importance weight of the corresponding channel. The time-dimensional broadcasting is an operation that replicates and expands the channel importance weight vector along the time dimension into a matrix with the same shape as the multi-scale temporal feature. The channel weight temporal matrix is ​​the matrix obtained after time-dimensional broadcasting. The channel-by-channel multiplication is an operation that multiplies the channel weight temporal matrix one by one with the elements of the corresponding channel and the corresponding time step in the multi-scale temporal feature.

[0055] For example, assuming a window length parameter of 30, 21 sensor channels, and a relative scale coefficient set containing three coefficients (0.1, 0.2, 0.4), taking the first relative scale coefficient of 0.1 as an example, the kernel size calculation process is as follows: 0.1 multiplied by 30 equals 3.0, rounded down to 3, 3 multiplied by 2 plus 1 equals 7, therefore the size of the first scale convolution kernel is 7. Similarly, the size of the second scale convolution kernel is 13, and the size of the third scale convolution kernel is 25. The modulation features are subjected to depthwise convolution operations using the three scale convolution kernels respectively, resulting in three sets of local single-channel convolution features, each with a shape of 30 x 21. Simultaneously, standard convolution operations are performed using the three scale convolution kernels respectively, resulting in three sets of global cross-channel convolution features, each with a shape of 30 x 21. The six sets of features are concatenated along the channel dimension to obtain a multi-scale temporal feature with a shape of 30 x 126. A time-dimensional adaptive average pooling operation is performed on the multi-scale temporal feature to obtain a 126-dimensional channel description vector. The channel description vector is input into a two-layer fully connected gated network with a dimensionality reduction ratio of 6. The first layer compresses the 126-dimensional channel to 21-dimensional channel and activates it using ReLU. The second layer restores the 126-dimensional channel and activates it using sigmoid, resulting in a 126-dimensional channel importance weight vector. This channel importance weight vector is then broadcast along the time dimension as a 30x126 channel weight temporal matrix, which is multiplied channel-by-channel with multi-scale temporal features to obtain a 30x126 aggregated degradation feature. This embodiment of the invention obtains degradation features at different temporal granularities and sensor association levels through multi-scale local and global convolutions, and adaptively adjusts the contribution of each channel using a channel gating mechanism, thereby improving the comprehensiveness and specificity of the degradation feature representation.

[0056] S5. Input the aggregated degradation features into the pre-constructed attention encoding branch and the pre-constructed state space encoding branch respectively to obtain the attention hidden representation and the state space hidden representation. Perform gating fusion on the attention hidden representation and the state space hidden representation to obtain the fused degradation hidden representation. Obtain the predicted value of the remaining service life of the device based on the fused degradation hidden representation.

[0057] It should be explained that the attention encoding branch is a deep encoding network based on a sparse self-attention mechanism, used to model long-distance temporal dependencies of aggregated degradation features and obtain global temporal correlation features of device degradation. Optionally, an Informer encoder can be used to achieve this purpose. The state space encoding branch is a deep encoding network based on a selective state space model, used to model long-sequence state propagation of aggregated degradation features with linear computational complexity and obtain continuous state evolution features of device degradation. Optionally, a selective state space model encoder can be used to achieve this purpose. The attention latent representation is the hidden feature representation obtained after processing the aggregated degradation features by the attention encoding branch, containing global attention relationship information. The state space latent representation is the hidden feature representation obtained after processing the aggregated degradation features by the state space encoding branch, containing continuous state propagation information. The gating fusion is an operation that uses learnable gating coefficients to adaptively weight and combine the attention latent representation and the state space latent representation. The fused degradation latent representation is a unified latent representation obtained after gating fusion of the attention latent representation and the state space latent representation, combining the global correlation modeling capability of the attention mechanism and the efficient sequence propagation capability of the state space model. The predicted remaining useful life of the equipment is based on the number of remaining operating cycles of the equipment from the current time to the time of failure, calculated using the fused degradation implicit representation.

[0058] Understandably, predicting equipment degradation trends requires simultaneously modeling long-distance time dependencies and continuous state evolution. A single encoding branch may not be able to fully capture the temporal information in the degradation features. Therefore, an attention encoding branch and a state space encoding branch are used in parallel, and adaptive feature integration is achieved through gating fusion. The aggregated degradation features are respectively input into the pre-constructed attention encoding branch and the pre-constructed state space encoding branch to obtain attention latent representations and state space latent representations. Gating fusion is then performed on the attention latent representations and state space latent representations to obtain a fused degradation latent representation, including: A relative time stamp sequence is generated based on the window length parameter. The aggregated degradation feature and the relative time stamp sequence are input into the pre-constructed feature case embedding mapping layer in the attention encoding branch to obtain the encoder input embedding. The encoder input is then embedded into the pre-constructed multi-layer sparse self-attention encoder to obtain the attention hidden representation. The aggregated degenerate features are input into the pre-constructed selective state space model in the state space coding branch to obtain the state space latent representation; The attention latent representation and the state space latent representation are gated and fused to obtain the fused degenerate latent representation. The calculation formula for the gated fusion is as follows:

[0059]

[0060] in, Indicates the gating fusion coefficient. This represents the preset fusion layer weight matrix. This represents the attention hidden representation. This represents the implicit representation of the state space. This represents the preset fusion layer bias vector. This represents the fusion-degenerate hidden representation. This indicates an element-wise multiplication operation.

[0061] In detail, the relative time stamp sequence is a sequence of normalized time position stamps generated for each acquisition time point in the sensor window, arranged in chronological order. Specifically, it is generated by dividing the index of each acquisition time point within the window by the window length parameter to obtain a relative time stamp with a value between zero and one. This time stamp only describes the relative time order within the window and does not introduce information about future acquisition time points. The feature condition embedding mapping layer is a network layer that jointly encodes aggregated degenerate features and the relative time stamp sequence into a unified embedding representation. It maps the aggregated degenerate features to the embedding dimension space through a linear transformation and superimposes the positional encoding information corresponding to the relative time stamps. The encoder input embedding is a feature representation with time position information obtained after the aggregated degenerate features have been processed by the feature condition embedding mapping layer. The described multilayer sparse self-attention encoder is a multilayer encoder structure. Its core idea is that in the attention calculation of each layer, instead of calculating the correlation scores between all pairs of acquisition time points, it first evaluates the dispersion of attention scores when each acquisition time point is used as a query. Then, it selects only acquisition time points with a large dispersion of attention scores for complete attention calculation. For acquisition time points with relatively uniform attention scores, it uses an approximation of the mean, thereby reducing computational load while maintaining long-range dependency modeling capabilities. The described selective state-space model is a sequence modeling structure. Its core idea is to maintain an internal state vector that updates with each acquisition time point. At each acquisition time point, the state vector of the previous acquisition time point is updated to the current acquisition time point's state vector using a state transition matrix (i.e., a parameter matrix that controls how the current state evolves from the state of the previous acquisition time point). Simultaneously, the input projection matrix (i.e., a parameter matrix that maps the input features of the current acquisition time point to the state space) is used to integrate the input features of the current acquisition time point into the state vector, and the state transition parameters are adaptively adjusted according to the input content, thereby achieving efficient modeling of long sequences.

[0062] Understandably, the fusion layer weight matrix and fusion layer bias vector are obtained as follows: the input dimension is determined based on the dimension of the concatenated attention latent representation and state space latent representation, and the output dimension is determined based on the latent representation dimension, thus determining the size of the fusion layer weight matrix. Then, the fusion layer weight matrix is ​​randomly initialized using a normal distribution with a mean of 0 and a standard deviation of 0.02. The fusion layer bias vector is initialized to zero. During end-to-end model training, the loss function is calculated, and the gradient is calculated using the backpropagation algorithm. Iterative optimization is performed using the Adam optimizer, and finally, the parameters after training convergence are saved. The gated fusion coefficient is a vector output by the sigmoid activation function after linearly transforming the concatenated vector of the attention latent representation and state space latent representation using the fusion layer weight matrix and fusion layer bias vector. Each element has a value ranging from 0 to 1 and is used to control the contribution ratio of the attention latent representation and the state space latent representation in the fusion of the degenerate latent representation.

[0063] It should be explained that, in order to obtain the predicted value of the remaining useful life of the device based on the fused degenerate implicit representation, it is necessary to map the implicit representation into a scalar prediction output through a decoder structure. The process of obtaining the predicted value of the remaining useful life of the device based on the fused degenerate implicit representation includes: Extract the terminal history fragment from the fused degenerate hidden representation, and concatenate the terminal history fragment with the preset prediction placeholder vector to obtain the decoder input sequence; The hidden state at the predicted location is obtained by performing cross-attention calculation on the decoder input sequence and the attention hidden representation using a pre-defined decoder. A linear projection operation is performed on the predicted hidden state of the location to obtain the predicted value of the remaining service life of the device.

[0064] It should be understood that the terminal historical fragment is the last continuous feature fragment extracted from the fused degenerate latent representation in the time dimension. It contains the device's recent degenerate latent layer information. The length of the terminal historical fragment can be set according to the decoder's label length parameter; optionally, the label length parameter can be set to half the window length parameter. The predicted placeholder vector is a zero-valued vector with the same dimension as the fused degenerate latent representation. It is used to identify the position to be predicted and does not contain any future real data information. The decoder input sequence is obtained by concatenating the terminal historical fragment and the predicted placeholder vector along the time dimension. The decoder is a multi-layer decoder structure based on an attention mechanism, which includes a self-attention layer and a cross-attention layer. The cross-attention calculation is a calculation process in which the decoder input sequence is used as a query, and the attention latent representation is used as the key and value. The recent historical degenerate features are aligned and fused with the global historical degenerate representation through the attention mechanism. The purpose is to guide the global degenerate trend information to the predicted position. The predicted position latent state is the latent feature representation of the position corresponding to the predicted placeholder vector in the decoder output. The linear projection operation is an operation that uses a pre-trained linear projection layer to map the predicted location hidden state from a high-dimensional feature space to a one-dimensional scalar space. The predicted remaining device lifetime is a scalar value output by the linear projection operation, representing the number of operating cycles the device is expected to continue running from the current moment.

[0065] For example, assume the shape of the aggregated degenerate feature is 30 x 126. A relative time-stamped sequence is generated based on the window length parameter 30, where the relative time-stamp of the first acquisition time point is 1 / 30 = 0.033, and the relative time-stamp of the 30th acquisition time point is 30 / 30 = 1.0. The aggregated degenerate feature and the relative time-stamped sequence are input into the feature embedding mapping layer to obtain the encoder input embedding. This encoder input is then embedded into a multi-layer sparse self-attention encoder containing three layers to obtain a 256-dimensional attention latent representation. Simultaneously, the aggregated degenerate feature is input into a selective state-space model to obtain a 256-dimensional state-space latent representation. The attention latent representation and the state-space latent representation are concatenated along the feature dimension to form a 512-dimensional vector, which is then linearly transformed into the fusion layer weight matrix (size 256 x 512) and the fusion layer bias vector. This transformation is followed by a sigmoid activation function to obtain 256-dimensional gated fusion coefficients. The 256-dimensional fused degenerate latent representation is then calculated according to the gated fusion formula. The label length parameter is set to 15. Features corresponding to the last 15 acquisition time points are extracted from the fused degradation latent representation as the final historical fragment. This final historical fragment is concatenated with a 256-dimensional all-zero prediction placeholder vector to obtain a decoder input sequence of length 16. A two-layer decoder is used to perform cross-attention calculation on the decoder input sequence and the attention latent representation to obtain the predicted position latent state. A linear projection operation is performed on the predicted position latent state to obtain a predicted remaining service life of 87.3 working cycles, indicating that the turbofan engine is expected to continue operating for approximately 87 more working cycles. This embodiment of the invention achieves accurate prediction of the remaining service life of the equipment by modeling in parallel with attention encoding branches and state space encoding branches, combining a gating fusion mechanism to adaptively integrate global attention relationships and continuous state transmission information, and using the decoder cross-attention structure to guide the global degradation trend to the predicted position.

[0066] To address the problems described in the background art, this invention acquires equipment operation monitoring data, identifies the operating condition setting sequence and multi-channel sensor numerical sequences based on the monitoring data, standardizes the multi-channel sensor numerical sequences to obtain standardized sensor numerical sequences, and constructs sensor windows and operating condition setting windows using the standardized sensor numerical sequences and operating condition setting sequences. It is evident that this invention eliminates differences in dimensions and numerical ranges between different sensor channels by calculating the mean and standard deviation for each channel, enabling the model to perform feature learning at a unified numerical scale. Based on the operating condition setting sequences, a set of operating condition cluster centers is constructed, and these cluster centers are used to analyze the operating conditions. The system assigns operating condition identifiers to a window to obtain an operating condition identifier sequence. Based on the operating condition identifier sequence, it obtains operating condition statistical nodes. Using the operating condition statistical nodes and the operating condition identifier sequence, it performs channel-level adaptive modulation on the sensor window to obtain operating condition modulation features. It can be seen that the present invention generates channel-level modulation signals based on the prefix statistics of the operating condition identifier sequence, enabling the model to obtain operating condition perception capabilities before extracting degradation features. It distinguishes the contributions of operating condition factors and degradation factors in sensor value changes. Local single-channel convolution and global cross-channel convolution are performed on the operating condition modulation features respectively, and the features are concatenated to obtain multi-scale temporal features. The multi-scale temporal features are then subjected to time-dimensional pooling and channel-gated weighting to obtain aggregated degradation features. As can be seen, this invention acquires degradation features at different temporal granularities and sensor association levels through multi-scale local and global convolutions, and adaptively adjusts the contribution of each channel using a channel gating mechanism, thereby improving the comprehensiveness and specificity of degradation feature representation. The aggregated degradation features are input into pre-constructed attention encoding branches and pre-constructed state space encoding branches to obtain attention latent representations and state space latent representations. These latent representations are then gated and fused to obtain a fused degradation latent representation. Based on this fused degradation latent representation, the predicted remaining lifespan of the device is obtained. Thus, this invention achieves accurate prediction of the remaining lifespan of the device by modeling in parallel with attention encoding branches and state space encoding branches, adaptively integrating global attention relationships and continuous state transmission information using a gating fusion mechanism, and guiding the global degradation trend to the prediction position using a decoder cross-attention structure. Therefore, this invention can achieve accurate prediction of remaining lifespan under multiple operating conditions.

[0067] like Figure 2 The diagram shown is a functional block diagram of a multi-condition life prediction system based on multimodal fusion provided in an embodiment of the present invention.

[0068] The multi-mode fusion-based multi-condition life prediction system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the multi-mode fusion-based multi-condition life prediction system 100 may include a data preprocessing module 101, a condition modulation module 102, a feature extraction module 103, and a life prediction module 104. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0069] The data preprocessing module 101 is used to acquire equipment operation monitoring data and determine the operating condition setting sequence and multi-channel sensor value sequence based on the equipment operation monitoring data. The multi-channel sensor numerical sequence is standardized to obtain a standardized sensor numerical sequence. The sensor window and the operating condition setting window are constructed using the standardized sensor numerical sequence and the operating condition setting sequence. The operating condition modulation module 102 is used to construct an operating condition cluster center set based on the operating condition setting sequence, use the operating condition cluster center set to allocate operating condition identifiers to the operating condition setting window to obtain an operating condition identifier sequence, obtain operating condition statistical nodes based on the operating condition identifier sequence, and use the operating condition statistical nodes and the operating condition identifier sequence to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features. The feature extraction module 103 is used to perform local single-channel convolution and global cross-channel convolution on the working condition modulation features respectively, and concatenate them to obtain multi-scale temporal features. The multi-scale temporal features are then subjected to time-dimensional pooling and channel-gated weighting to obtain aggregated degradation features. The lifetime prediction module 104 is used to input the aggregated degradation features into a pre-constructed attention encoding branch and a pre-constructed state space encoding branch to obtain attention hidden representation and state space hidden representation, perform gating fusion on the attention hidden representation and state space hidden representation to obtain fused degradation hidden representation, and obtain the predicted value of the remaining lifetime of the device based on the fused degradation hidden representation.

[0070] In detail, the modules in the multi-mode fusion-based multi-condition lifetime prediction system 100 described in this embodiment of the invention employ the same methods as described above. Figure 1 The method used here is the same as the multi-condition life prediction method based on multi-modal fusion described above, and can produce the same technical effect, so it will not be repeated here.

[0071] like Figure 3 The diagram shown is a structural schematic of an electronic device that implements a multi-condition lifetime prediction method based on multi-modal fusion, according to an embodiment of the present invention.

[0072] The electronic device 1 may include a processor 10, a memory 11 and a bus 12, and may also include a computer program stored in the memory 11 and capable of running on the processor 10, such as a multi-condition lifetime prediction method program based on multimodal fusion.

[0073] The memory 11 includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of the electronic device 1, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device 1, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 1. Furthermore, the memory 11 includes both internal storage units and external storage devices of the electronic device 1. The memory 11 can be used not only to store application software and various types of data installed on the electronic device 1, such as the code of a multi-mode fusion-based multi-condition life prediction method program, but also to temporarily store data that has been output or will be output.

[0074] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 (e.g., a multi-modal fusion-based multi-condition lifetime prediction method program), and calls data stored in the memory 11 to perform various functions of the electronic device 1 and process data.

[0075] The bus 12 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus 12 can be divided into an address bus, a data bus, a control bus, etc. The bus 12 is configured to realize the connection and communication between the memory 11 and at least one processor 10, etc.

[0076] Figure 3 Only electronic devices with components are shown; it will be understood by those skilled in the art that... Figure 3 The structure shown does not constitute a limitation on the electronic device 1, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0077] For example, although not shown, the electronic device 1 may also include a power supply (such as a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 10 through a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device 1 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0078] Furthermore, the electronic device 1 may also include a network interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, etc.), which is typically used to establish communication connections between the electronic device 1 and other electronic devices.

[0079] Optionally, the electronic device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), and optionally, a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device 1 and to display a visual user interface.

[0080] The program for a multi-condition lifetime prediction method based on multimodal fusion, stored in the memory 11 of the electronic device 1, is a combination of multiple instructions. When run in the processor 10, it can achieve the following: Acquire equipment operation monitoring data, and based on the equipment operation monitoring data, determine the operating condition setting sequence and the multi-channel sensor value sequence; The multi-channel sensor numerical sequence is standardized to obtain a standardized sensor numerical sequence. The sensor window and the operating condition setting window are constructed using the standardized sensor numerical sequence and the operating condition setting sequence. A set of operating condition cluster centers is constructed based on the operating condition setting sequence. The operating condition cluster centers are used to assign operating condition identifiers to the operating condition setting window to obtain an operating condition identifier sequence. Operating condition statistical nodes are obtained based on the operating condition identifier sequence. The operating condition statistical nodes and the operating condition identifier sequence are used to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features. Local single-channel convolution and global cross-channel convolution are performed on the aforementioned operating condition modulation features respectively, and the features are concatenated to obtain multi-scale temporal features. Temporal pooling and channel-gated weighting are then applied to the multi-scale temporal features to obtain aggregated degradation features. The aggregated degradation features are input into the pre-constructed attention encoding branch and the pre-constructed state space encoding branch to obtain the attention hidden representation and the state space hidden representation. The attention hidden representation and the state space hidden representation are then gated and fused to obtain the fused degradation hidden representation. Based on the fused degradation hidden representation, the predicted value of the remaining service life of the device is obtained.

[0081] Specifically, the processor 10's implementation method for the above instructions can be found in [reference needed]. Figures 1 to 3 The descriptions of the relevant steps in the corresponding embodiments are not repeated here.

[0082] Furthermore, if the modules / units integrated in the electronic device 1 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0083] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor of an electronic device, can perform the following: Acquire equipment operation monitoring data, and based on the equipment operation monitoring data, determine the operating condition setting sequence and the multi-channel sensor value sequence; The multi-channel sensor numerical sequence is standardized to obtain a standardized sensor numerical sequence. The sensor window and the operating condition setting window are constructed using the standardized sensor numerical sequence and the operating condition setting sequence. A set of operating condition cluster centers is constructed based on the operating condition setting sequence. The operating condition cluster centers are used to assign operating condition identifiers to the operating condition setting window to obtain an operating condition identifier sequence. Operating condition statistical nodes are obtained based on the operating condition identifier sequence. The operating condition statistical nodes and the operating condition identifier sequence are used to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features. Local single-channel convolution and global cross-channel convolution are performed on the aforementioned operating condition modulation features respectively, and the features are concatenated to obtain multi-scale temporal features. Temporal pooling and channel-gated weighting are then applied to the multi-scale temporal features to obtain aggregated degradation features. The aggregated degradation features are input into the pre-constructed attention encoding branch and the pre-constructed state space encoding branch to obtain the attention hidden representation and the state space hidden representation. The attention hidden representation and the state space hidden representation are then gated and fused to obtain the fused degradation hidden representation. Based on the fused degradation hidden representation, the predicted value of the remaining service life of the device is obtained.

[0084] In the embodiments provided by this invention, it should be understood that the disclosed devices, systems, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and actual implementations may have other classification methods.

[0085] The modules described as separate components may or may not be physically separate. The components shown as modules 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 embodiment according to actual needs.

[0086] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0087] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-condition lifetime prediction method based on multimodal fusion, characterized in that, The method includes: Acquire equipment operation monitoring data, and based on the equipment operation monitoring data, determine the operating condition setting sequence and the multi-channel sensor value sequence; The multi-channel sensor numerical sequence is standardized to obtain a standardized sensor numerical sequence. The sensor window and the operating condition setting window are constructed using the standardized sensor numerical sequence and the operating condition setting sequence. The standardization process of the multi-channel sensor numerical sequence to obtain a standardized sensor numerical sequence, and the construction of a sensor window and a working condition setting window using the standardized sensor numerical sequence and the operating condition setting sequence, includes: Multiple sensor channels and sensor value sets were identified based on multi-channel sensor numerical sequences; The multi-channel sensor numerical sequence is divided according to multiple sensor channels to obtain multiple channel numerical sequences; Perform the following operation on each channel of the multi-channel numerical sequence: Calculate the mean and standard deviation of the channel numerical sequences respectively to obtain the channel mean and channel standard deviation; By correlating the channel mean and channel standard deviation, channel standardized parameter nodes are obtained. By summarizing the channel standardized parameter nodes, a set of channel standardized parameter nodes is obtained, wherein each channel standardized parameter node corresponds one-to-one with a sensor channel. The sensor value set is standardized by using the channel standardized parameter node set, and then divided according to multiple sensor channels to obtain multiple standardized channel value sequences. According to the channel order of multiple sensor channels in the multi-channel sensor numerical sequence, the multiple standardized channel numerical sequences are sorted to obtain the standardized sensor numerical sequence. Obtain the window length parameter, and based on the window length parameter, slide and extract continuous segments in the standardized sensor value sequence and the operating condition setting sequence to obtain the sensor window and the operating condition setting window. A set of operating condition cluster centers is constructed based on the operating condition setting sequence. The operating condition cluster centers are used to assign operating condition identifiers to the operating condition setting window to obtain an operating condition identifier sequence. Operating condition statistical nodes are obtained based on the operating condition identifier sequence. The operating condition statistical nodes and the operating condition identifier sequence are used to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features. The step of obtaining operating condition statistical nodes based on the operating condition identifier sequence, and using the operating condition statistical nodes and the operating condition identifier sequence to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features includes: Obtain the latest time point and the latest time point sequence number from the equipment operation monitoring data; Input the time point working condition identifier corresponding to the latest time point into the pre-constructed working condition embedding mapping layer to obtain the working condition embedding vector. Input the working condition embedding vector into the pre-constructed multilayer sensing network and perform activation processing to obtain the working condition projection vector. The total number of times the operating condition identifier changes between adjacent time points in the statistical operating condition identifier sequence is obtained as the number of operating condition switching. The ratio of the number of operating condition switching to the latest time point sequence number is calculated to obtain the operating condition switching frequency. The number of times the working condition identifier corresponding to the latest time point has appeared in the working condition identifier sequence is counted to obtain the cumulative number of working conditions. The ratio of the cumulative number of working conditions to the latest time point sequence number is calculated to obtain the window working condition percentage. By associating the window condition ratio and condition switching frequency, a condition statistics node is obtained. The condition statistics node is then input into a pre-constructed learnable gating layer to obtain a channel modulation intensity vector. The channel modulation bias is obtained by multiplying the channel modulation intensity vector with the operating condition projection vector. The sensor vector at the latest time point in the sensor window is added to the channel modulation bias to obtain the modulation sensor vector at the latest time point. By summarizing the modulation sensor vectors, the modulation characteristics under operating conditions can be obtained; The operating condition statistics node is input into a pre-constructed learnable gating layer to obtain the channel modulation intensity vector, and the calculation formula is as follows: in, This represents the channel modulation intensity vector. This indicates that the pre-built sigmoid activation function in the learnable gated layer is used. This indicates the pre-defined gate weight matrix in the learnable gate layer. This indicates the percentage of window conditions at the latest time point. This indicates the frequency of operating condition switching at the latest point in time. This represents the preset gate bias vector in the learnable gate layer. This represents a vector concatenation operation; Local single-channel convolution and global cross-channel convolution are performed on the aforementioned operating condition modulation features respectively, and the features are concatenated to obtain multi-scale temporal features. Temporal pooling and channel-gated weighting are then applied to the multi-scale temporal features to obtain aggregated degradation features. The aggregated degradation features are respectively input into the pre-constructed attention encoding branch and the pre-constructed state space encoding branch to obtain the attention hidden representation and the state space hidden representation. The attention hidden representation and the state space hidden representation are gated and fused to obtain the fused degradation hidden representation. The predicted value of the remaining service life of the device is obtained based on the fused degradation hidden representation. The process involves inputting the aggregated degenerate features into a pre-constructed attention encoding branch and a pre-constructed state space encoding branch to obtain an attention latent representation and a state space latent representation, and then performing gated fusion on the attention latent representation and the state space latent representation to obtain a fused degenerate latent representation, including: A relative time stamp sequence is generated based on the window length parameter. The aggregated degradation feature and the relative time stamp sequence are input into the pre-constructed feature case embedding mapping layer in the attention encoding branch to obtain the encoder input embedding. The encoder input is then embedded into the pre-constructed multi-layer sparse self-attention encoder to obtain the attention hidden representation. The aggregated degenerate features are input into the pre-constructed selective state space model in the state space coding branch to obtain the state space latent representation; The attention latent representation and the state space latent representation are gated and fused to obtain the fused degenerate latent representation. The calculation formula for the gated fusion is as follows: in, Indicates the gating fusion coefficient. This represents the preset fusion layer weight matrix. This represents the attention hidden representation. This represents the implicit representation of the state space. This represents the preset fusion layer bias vector. This represents the fusion-degenerate hidden representation. This indicates an element-wise multiplication operation.

2. The multi-condition lifetime prediction method based on multi-modal fusion as described in claim 1, characterized in that, The step involves constructing a set of operating condition cluster centers based on the operating condition setting sequence, and then using this set of cluster centers to assign operating condition identifiers to the operating condition setting window, resulting in an operating condition identifier sequence, including: Multiple time points are obtained from the collection of equipment operation monitoring data, resulting in multiple collection time points; Obtain working condition setting vectors corresponding to multiple collection time points to obtain a training working condition vector set. Perform mean clustering on the training working condition vector set to obtain a working condition cluster center set, wherein the working condition cluster center set includes multiple working condition cluster centers. In the operating condition setting window, extract the window operating condition setting vectors sequentially according to the chronological order of the data acquisition time points, and perform the following operations on the extracted window operating condition setting vectors: Calculate the vector distance between the working condition setting vector and each working condition cluster center in the working condition cluster center set to obtain a vector distance set, which includes multiple vector distances; The working condition cluster center corresponding to the smallest vector distance in the vector distance set is identified to obtain the target working condition center. Based on the target working condition center, the working condition identifier is assigned to the collection time point corresponding to the working condition setting vector to obtain the time point working condition identifier. By summarizing the operating condition identifiers at the aforementioned time points, a sequence of operating condition identifiers is obtained.

3. The multi-condition lifetime prediction method based on multi-modal fusion as described in claim 2, characterized in that, The process involves performing local single-channel convolution and global cross-channel convolution on the modulated features of the operating conditions, and then concatenating them to obtain multi-scale temporal features, including: Obtain a set of relative scale coefficients, wherein the set of relative scale coefficients includes multiple relative scale coefficients, and generate multiple scale convolution kernels based on the set of relative scale coefficients and the window length parameter; For each of the multiple scale convolution kernels, perform the following operation: The scale convolution kernel is used to perform channel-by-channel independent depth convolution operation on the working condition modulation feature to obtain local single-channel convolution feature, and the scale convolution kernel is used to perform cross-channel standard convolution operation on the working condition modulation feature to obtain global cross-channel convolution feature. By summarizing local single-channel convolutional features and global cross-channel convolutional features and performing dimensional concatenation, multi-scale temporal features are obtained.

4. The multi-condition lifetime prediction method based on multi-modal fusion as described in claim 3, characterized in that, The process of performing time-dimensional pooling and channel-gated weighting on the multi-scale temporal features to obtain aggregated degenerate features includes: Perform time-dimensional adaptive average pooling on multi-scale temporal features to obtain channel description vectors; The channel description vector is input into a pre-constructed two-layer fully connected gated network to obtain the channel importance weight vector; The channel importance weight vector is broadcast over time to obtain the channel weight time series matrix. The channel weight time series matrix is ​​then multiplied channel by channel by multi-scale time series features to obtain aggregated degradation features.

5. The multi-condition lifetime prediction method based on multi-modal fusion as described in claim 4, characterized in that, The step of obtaining the predicted remaining useful life of the device based on the fused degradation latent representation includes: Extract the terminal history fragment from the fused degenerate hidden representation, and concatenate the terminal history fragment with the preset prediction placeholder vector to obtain the decoder input sequence; The hidden state at the predicted location is obtained by performing cross-attention calculation on the decoder input sequence and the attention hidden representation using a pre-defined decoder. A linear projection operation is performed on the predicted hidden state of the location to obtain the predicted value of the remaining service life of the device.

6. A multi-condition life prediction system based on multimodal fusion, characterized in that, The system includes: The data preprocessing module is used to acquire equipment operation monitoring data and, based on the equipment operation monitoring data, determine the operating condition setting sequence and the multi-channel sensor value sequence. The multi-channel sensor numerical sequence is standardized to obtain a standardized sensor numerical sequence. The sensor window and the operating condition setting window are constructed using the standardized sensor numerical sequence and the operating condition setting sequence. The standardization process of the multi-channel sensor numerical sequence to obtain a standardized sensor numerical sequence, and the construction of a sensor window and a working condition setting window using the standardized sensor numerical sequence and the operating condition setting sequence, includes: Multiple sensor channels and sensor value sets were identified based on multi-channel sensor numerical sequences; The multi-channel sensor numerical sequence is divided according to multiple sensor channels to obtain multiple channel numerical sequences; Perform the following operation on each channel of the multi-channel numerical sequence: Calculate the mean and standard deviation of the channel numerical sequences respectively to obtain the channel mean and channel standard deviation; By correlating the channel mean and channel standard deviation, channel standardized parameter nodes are obtained. By summarizing the channel standardized parameter nodes, a set of channel standardized parameter nodes is obtained, wherein each channel standardized parameter node corresponds one-to-one with a sensor channel. The sensor value set is standardized by using the channel standardized parameter node set, and then divided according to multiple sensor channels to obtain multiple standardized channel value sequences. According to the channel order of multiple sensor channels in the multi-channel sensor numerical sequence, the multiple standardized channel numerical sequences are sorted to obtain the standardized sensor numerical sequence. Obtain the window length parameter, and based on the window length parameter, slide and extract continuous segments in the standardized sensor value sequence and the operating condition setting sequence to obtain the sensor window and the operating condition setting window. The operating condition modulation module is used to construct an operating condition cluster center set based on the operating condition setting sequence, use the operating condition cluster center set to allocate operating condition identifiers to the operating condition setting window to obtain an operating condition identifier sequence, obtain operating condition statistical nodes based on the operating condition identifier sequence, and use the operating condition statistical nodes and the operating condition identifier sequence to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features. The step of obtaining operating condition statistical nodes based on the operating condition identifier sequence, and using the operating condition statistical nodes and the operating condition identifier sequence to perform channel-level adaptive modulation on the sensor window to obtain operating condition modulation features includes: Obtain the latest time point and the latest time point sequence number from the equipment operation monitoring data; Input the time point working condition identifier corresponding to the latest time point into the pre-constructed working condition embedding mapping layer to obtain the working condition embedding vector. Input the working condition embedding vector into the pre-constructed multilayer sensing network and perform activation processing to obtain the working condition projection vector. The total number of times the operating condition identifier changes between adjacent time points in the statistical operating condition identifier sequence is obtained as the number of operating condition switching. The ratio of the number of operating condition switching to the latest time point sequence number is calculated to obtain the operating condition switching frequency. The number of times the working condition identifier corresponding to the latest time point has appeared in the working condition identifier sequence is counted to obtain the cumulative number of working conditions. The ratio of the cumulative number of working conditions to the latest time point sequence number is calculated to obtain the window working condition percentage. By associating the window condition ratio and condition switching frequency, a condition statistics node is obtained. The condition statistics node is then input into a pre-constructed learnable gating layer to obtain a channel modulation intensity vector. The channel modulation bias is obtained by multiplying the channel modulation intensity vector with the operating condition projection vector. The sensor vector at the latest time point in the sensor window is added to the channel modulation bias to obtain the modulation sensor vector at the latest time point. By summarizing the modulation sensor vectors, the modulation characteristics under operating conditions can be obtained; The operating condition statistics node is input into a pre-constructed learnable gating layer to obtain the channel modulation intensity vector, and the calculation formula is as follows: in, This represents the channel modulation intensity vector. This indicates that the pre-built sigmoid activation function in the learnable gated layer is used. This indicates the pre-defined gate weight matrix in the learnable gate layer. This indicates the percentage of window conditions at the latest time point. This indicates the frequency of operating condition switching at the latest point in time. This represents the preset gate bias vector in the learnable gate layer. This represents a vector concatenation operation; The feature extraction module is used to perform local single-channel convolution and global cross-channel convolution on the modulated features of the working conditions respectively, and concatenate them to obtain multi-scale temporal features. The multi-scale temporal features are then subjected to temporal pooling and channel-gated weighting to obtain aggregated degradation features. The lifetime prediction module is used to input the aggregated degradation features into a pre-constructed attention encoding branch and a pre-constructed state space encoding branch to obtain attention hidden representation and state space hidden representation, perform gating fusion on the attention hidden representation and state space hidden representation to obtain fused degradation hidden representation, and obtain the predicted value of the remaining lifetime of the device based on the fused degradation hidden representation. The process involves inputting the aggregated degenerate features into a pre-constructed attention encoding branch and a pre-constructed state space encoding branch to obtain an attention latent representation and a state space latent representation, and then performing gated fusion on the attention latent representation and the state space latent representation to obtain a fused degenerate latent representation, including: A relative time stamp sequence is generated based on the window length parameter. The aggregated degradation feature and the relative time stamp sequence are input into the pre-constructed feature case embedding mapping layer in the attention encoding branch to obtain the encoder input embedding. The encoder input is then embedded into the pre-constructed multi-layer sparse self-attention encoder to obtain the attention hidden representation. The aggregated degenerate features are input into the pre-constructed selective state space model in the state space coding branch to obtain the state space latent representation; The attention latent representation and the state space latent representation are gated and fused to obtain the fused degenerate latent representation. The calculation formula for the gated fusion is as follows: in, Indicates the gating fusion coefficient. This represents the preset fusion layer weight matrix. This represents the attention hidden representation. This represents the implicit representation of the state space. This represents the preset fusion layer bias vector. This represents the fusion-degenerate hidden representation. This indicates an element-wise multiplication operation.

Images