Method and device for predicting service life of rotating machine based on multi-sensor spatio-temporal information fusion network

Abstract

The invention discloses a rotating machine life prediction method and device based on a multi-sensor spatio-temporal information fusion network, and relates to the technical field of rotating machine health state assessment, and the method comprises the steps: obtaining multi-source monitoring signal data of a to-be-predicted rotating machine in a set time period; performing normalization processing on the multi-source monitoring signal data; constructing a monitoring data sample based on a set sliding step length and the normalized multi-source monitoring signal data by adopting a set sliding window; constructing a time information mining module, a spatial relationship modeling module, a residual connection module and a full connection network, and performing program optimization processing to obtain a life prediction model; and inputting the monitoring data sample into the life prediction model to obtain the residual service life value of the to-be-predicted rotating machine. According to the invention, the monitoring signals collected by the multi-source sensor can be fused, the time information and the spatial relationship in the multi-source monitoring signals are extracted and fused, and the prediction precision and robustness of the life prediction model are improved.

Description

Technical Field

[0001] The present application relates to the technical field of rotating machinery health status assessment, and in particular to a rotating machinery life prediction method and device based on a multi-sensor spatiotemporal information fusion network. Background Art

[0002] Rotating machinery is widely used in production and daily life, covering fields such as wind power generation, aerospace, rail transportation, and industrial manufacturing. Because rotating machinery often operates in harsh environments such as variable speeds, high temperatures, and heavy loads, it is prone to failures such as wear and breakage, which can reduce equipment accuracy and reliability. In severe cases, this can lead to production interruptions and even casualties. Therefore, research on the remaining useful life prediction of rotating machinery is of great significance. By establishing a scientific life prediction model, it can provide information support for equipment maintenance decisions, optimize spare parts management, avoid resource waste, effectively prevent major accidents, and ensure the stable and efficient operation of industrial production systems.

[0003] With the accelerated iteration and industrialization of IoT technology, the deployment costs of various sensor types have significantly decreased, providing a critical technical foundation for the systematic application of multi-source sensors in rotating machinery condition monitoring. Compared to single-source sensor data, multi-source sensor data can provide more comprehensive and richer information on equipment degradation, potentially enabling more accurate remaining useful life prediction. However, the complex spatiotemporal coupling and inherent heterogeneity of multi-source sensor signals pose new challenges for life prediction. To address this issue, life prediction methods based on graph neural networks have emerged. However, existing approaches struggle to capture the non-stationary, time-varying characteristics of each sensor signal during degradation, resulting in insufficient accuracy in modeling dynamic temporal patterns. Furthermore, predefined static graph structures cannot represent the spatiotemporal relationships between sensor nodes that evolve dynamically with equipment degradation, while existing dynamic graph methods still have limitations in long-term modeling and prior knowledge integration. Summary of the Invention

[0004] The purpose of this application is to provide a rotating machinery life prediction method and equipment based on a multi-sensor spatiotemporal information fusion network, which can fuse the monitoring signals collected by multi-source sensors, extract and fuse the time information and spatial relationships in the multi-source monitoring signals, and improve the prediction accuracy and robustness of the life prediction model.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network, comprising:

[0007] Acquire multi-source monitoring signal data within a set period of time of the rotating machinery to be predicted; the multi-source monitoring signal data includes a monitoring signal sequence collected by multiple different types of sensors;

[0008] performing normalization processing on the multi-source monitoring signal data;

[0009] A sliding window of a set size is used to slide on the normalized multi-source monitoring signal data according to a set step size to construct a monitoring data sample;

[0010] Construct a temporal information mining module, a spatial relationship modeling module, a residual connection module, and a fully connected network;

[0011] Performing program optimization processing on the time information mining module, the spatial relationship modeling module, the residual connection module and the fully connected network to obtain a life prediction model;

[0012] The monitoring data sample is input into the life prediction model to obtain the remaining service life value of the rotating machinery to be predicted; wherein, the time information mining module is used to extract the cross-sequence relationship between the multi-source monitoring signal data in the monitoring data sample and the intra-sequence relationship of each monitoring signal sequence, and aggregate the cross-sequence relationship and the intra-sequence relationship to obtain the time characteristics of the monitoring data; the spatial relationship modeling module is used to obtain the spatiotemporal fusion characteristics of the monitoring data based on the time characteristics of the monitoring data; the residual module is used to perform residual connection on the monitoring data sample and the spatiotemporal fusion characteristics of the monitoring data to obtain the connected monitoring data; the fully connected network is used to obtain the remaining service life prediction value of the rotating machinery to be predicted based on the connected monitoring data.

[0013] Optionally, the time information mining module, the spatial relationship modeling module, the residual module and the fully connected network are subjected to program optimization processing to obtain a life prediction model, including:

[0014] Acquire multi-source monitoring signal data of the rotating machinery throughout its life cycle, and perform life annotation on the multi-source monitoring signal data of the rotating machinery throughout its life cycle to obtain training data;

[0015] Normalizing the training data to obtain normalized training data;

[0016] Using a sliding window of a set size, sliding on the normalized training data according to the set step size to construct a training data sample;

[0017] Programming the time information mining module, the spatial relationship modeling module, the residual module, and the fully connected network to obtain an initial prediction model;

[0018] The number of training rounds is set, and the initial prediction model is trained based on the training data samples. When the number of training rounds reaches the set number of training rounds, the training is stopped to obtain an optimized initial prediction model; and the optimized initial prediction model is used as the life prediction model.

[0019] Optionally, a self-attention mechanism is used to extract cross-sequence relationships between multi-source monitoring signal data in the monitoring data samples.

[0020] Optionally, in the time information mining module, the process of extracting the intra-sequence relationship of each monitoring signal sequence in the monitoring data sample includes:

[0021] Obtaining a time scale of a monitoring data sample based on the monitoring data sample using a fast Fourier transform;

[0022] A bidirectional Token sequence construction method is used to obtain a Token sequence of the monitoring data sample based on the time scale of the monitoring data sample;

[0023] The intra-sequence relationship of each monitoring signal data in the monitoring data sample is obtained based on the monitoring data sample and the Token sequence of the monitoring data sample.

[0024] Optionally, the spatial relationship modeling module includes a graph structure construction unit and a multi-channel graph convolutional network;

[0025] The graph structure construction unit is used to obtain a first adjacency matrix, a second adjacency matrix and a third adjacency matrix based on the time characteristics of the monitoring data;

[0026] The multi-channel graph convolutional network is used to obtain the spatiotemporal fusion features of the monitoring data based on the first adjacency matrix, the second adjacency matrix and the third adjacency matrix.

[0027] Optionally, in the graph structure construction unit, the process of obtaining the first adjacency matrix, the second adjacency matrix, and the third adjacency matrix based on the time characteristics of the monitoring data includes:

[0028] Determining the cosine similarity between each two monitoring signal sequences in the time characteristics of the monitoring data, and generating the first adjacency matrix based on the obtained cosine similarity values;

[0029] generating a second adjacency matrix based on a sensor type of the monitoring signal sequence corresponding to the time feature of the monitoring data;

[0030] The adjacency matrix is ​​generated based on the time information of the monitoring signal sequence in the monitoring data time feature.

[0031] Optionally, training the initial prediction model based on the training data sample, stopping the training when the number of training rounds reaches the set training round value, and using the trained initial prediction model as the life prediction model, includes:

[0032] Using the initial prediction model, based on the training data samples, a remaining life prediction value corresponding to the full life cycle signal data of the rotating machinery is obtained;

[0033] Optimizing the parameters of the initial prediction model by using the mean square error between the remaining life prediction value and the corresponding true life label value as an optimization function;

[0034] When the number of training rounds reaches the set training round value, the training is stopped to obtain an optimized initial prediction model; the optimized initial prediction model is used as the life prediction model.

[0035] In the second aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any one of the above-mentioned methods for predicting the life of rotating machinery based on a multi-sensor spatiotemporal information fusion network.

[0036] Optionally, the memory is a computer-readable storage medium.

[0037] In a third aspect, the present application provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of any one of the above-mentioned methods for predicting the life of rotating machinery based on a multi-sensor spatiotemporal information fusion network.

[0038] According to the specific embodiments provided in this application, this application has the following technical effects:

[0039] The present application provides a method, device, medium and product for predicting the life of rotating machinery based on a multi-sensor spatiotemporal information fusion network, which uses multiple different types of sensors to collect multi-source signal monitoring data of the rotating machinery to be predicted, processes it to obtain monitoring data samples, and processes the monitoring data samples through the constructed time information mining module, spatial relationship modeling module, residual connection module and fully connected network to obtain the remaining service life prediction value of the rotating machinery to be predicted. The time information mining module extracts the cross-sequence relationship between the multi-source monitoring signal data in the monitoring data sample and the intra-sequence relationship of each monitoring signal sequence, and aggregates the cross-sequence relationship and the intra-sequence relationship to obtain the time feature of the monitoring data, extracts the complex relationship between different sensor signals and the unsteady change characteristics during the degradation process of each sensor signal. And through the spatial relationship modeling module, the spatial relationship in the multi-source monitoring signal is extracted, and the dynamic time features and spatial relationships in the multi-source monitoring signal are fused, thereby improving the prediction accuracy and robustness of the life prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0041] Figure 1 This is a flow chart of a method for predicting the life of rotating machinery based on a multi-sensor spatiotemporal information fusion network in one embodiment of the present application;

[0042] Figure 2 A schematic diagram of a sliding window operation provided in an embodiment of the present application;

[0043] Figure 3 A schematic diagram of the structure of a time information mining module provided in one embodiment of the present application;

[0044] Figure 4 A schematic diagram of the spatial information modeling module structure provided in one embodiment of the present application;

[0045] Figure 5 A flowchart of a rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network according to an embodiment of the present application;

[0046] Figure 6 A schematic diagram of the engine structure of the engine used in the CMAPSS dataset provided in one embodiment of the present application;

[0047] Figure 7 A priori map constructed based on sensor type in the CMAPSS dataset provided in one embodiment of the present application;

[0048] Figure 8 A schematic diagram of the life prediction results of four test engines in the CMAPSS dataset provided in one embodiment of the present application;

[0049] Figure 9 A schematic diagram of the lifespan prediction results of FD001 in the CMAPSS dataset provided in one embodiment of the present application;

[0050] Figure 10 A schematic diagram of the lifespan prediction results of FD002 in the CMAPSS dataset provided in one embodiment of the present application;

[0051] Figure 11 A schematic diagram of the lifespan prediction results of FD003 in the CMAPSS dataset provided in one embodiment of the present application;

[0052] Figure 12 A schematic diagram of the lifespan prediction results of FD004 in the CMAPSS dataset provided in one embodiment of the present application;

[0053] Figure 13 A schematic diagram of sensor locations in SCADA provided in another embodiment of the present application;

[0054] Figure 14 A schematic diagram of a priori graph structure based on sensor type in SCADA provided by another embodiment of the present application;

[0055] Figure 15 A schematic diagram of the remaining service life prediction results of wind turbine No. 15 provided in another embodiment of the present application;

[0056] Figure 16 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0058] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0059] In an exemplary embodiment, Figure 1As shown, a method for predicting the life of rotating machinery based on a multi-sensor spatiotemporal information fusion network is provided. The method is executed by a computer device, specifically, a computer device such as a terminal or a server, or a terminal and a server. In the embodiment of the present application, the method is used to predict the remaining service life of rotating machinery as an example, and includes:

[0060] Step 100: Acquire multi-source monitoring signal data within a set period of time of the rotating machinery to be predicted. The multi-source monitoring signal data includes monitoring signal sequences collected by multiple different types of sensors.

[0061] Step 200: normalize the multi-source monitoring signal data.

[0062] Step 300 : Using a sliding window of a set size, sliding on the normalized multi-source monitoring signal data according to a set step size to construct a monitoring data sample.

[0063] Step 400: construct a time information mining module, a spatial relationship modeling module, a residual connection module and a fully connected network, and optimize the time information mining module, the spatial relationship modeling module, the residual connection module and the fully connected network to obtain a life prediction model.

[0064] In step 500, the monitoring data samples are input into the life prediction model to obtain the remaining service life value of the rotating machinery to be predicted. Among them, the time information mining module is used to extract the cross-sequence relationship between the multi-source monitoring signal data in the monitoring data sample and the intra-sequence relationship of each monitoring signal sequence, and aggregate the cross-sequence relationship and the intra-sequence relationship to obtain the time characteristics of the monitoring data. The spatial relationship modeling module is used to obtain the spatiotemporal fusion characteristics of the monitoring data based on the time characteristics of the monitoring data. The residual module is used to perform residual connection on the monitoring data samples and the spatiotemporal fusion characteristics of the monitoring data to obtain the connected monitoring data. The fully connected network is used to obtain the predicted value of the remaining service life of the rotating machinery to be predicted based on the connected monitoring data.

[0065] As an optional implementation, in order to improve the accuracy of the prediction result, step 400 includes:

[0066] Step 410: Acquire multi-source monitoring signal data of the rotating machinery throughout its life cycle, and perform life annotation on the multi-source monitoring signal data of the rotating machinery throughout its life cycle to obtain training data. For example, the multi-source monitoring signal data of the rotating machinery throughout its life cycle is collected by a multi-source sensor system. S=(x1,x2,...,x N ), x j =(x 1,j ,x 2,j ,...,x T,j ), x j represents the monitoring signal sequence of the entire life cycle obtained by the jth sensor, x t,j represents the jth monitoring signal sequence (i.e., the signal data collected by the jth sensor) in the tth sample (i.e., at time t). A multi-source sensor system is composed of multiple sensors of different types, N is the number of sensors, and T is the number of samples over the entire life cycle (in practical applications, T represents the number of multi-source monitoring signal data acquired within a set time period). The multi-source monitoring signal data over the entire life cycle of the rotating machinery is then labeled with corresponding lifespan to obtain training data.

[0067] Step 420: normalize the training data to obtain normalized training data. Use the set sliding window to construct training data samples based on the set sliding step size and the normalized training data. For example, use the formula Normalization processing is performed on multi-source monitoring signal data of the rotating machinery throughout its life cycle in the training data to obtain normalized training data. represents the jth monitoring signal sequence in the tth sample after normalization (i.e. at time t), represents the maximum value of the jth monitoring signal sequence, Represents the minimum value of the jth monitoring signal sequence. The sliding window operation is used to construct the training data sample based on the training data. The sliding window length is W, the sliding step is S, and each window sample is (i.e., the training data sample of the w-th sliding window, which represents the monitoring data sample of the w-th sliding window in practical applications), the obtained training data sample is The RUL of the w+1th window sample (i.e., sliding window) is TWw*S, such as Figure 2 shown.

[0068] It should be noted that in the early operation stage of rotating machinery, the degree of equipment degradation is small and it can be assumed to be in a healthy state. The remaining service life in this stage is constant. After running for a period of time, it enters the linear degradation stage. Therefore, the remaining service life y in these two stages is * Defined as: Where RUL max Indicates the maximum remaining useful life, and RUL stands for Remaining Useful Life (RUL).

[0069] Step 430 , program the temporal information mining module, the spatial relationship modeling module, the residual module, and the fully connected network to obtain an initial prediction model.

[0070] Step 440: Set a number of training rounds, train the initial prediction model based on the training data samples, stop training when the number of training rounds reaches the set number of training rounds, and obtain an optimized initial prediction model. The optimized initial prediction model is used as the lifespan prediction model.

[0071] Furthermore, in step 440, the initial prediction model is trained based on the training data samples, and the training is stopped when the number of training rounds reaches the set number of training rounds to obtain an optimized initial prediction model, which is used as the life prediction model, including:

[0072] Step 441 : Using the initial prediction model, based on the training data samples, obtain the remaining life prediction value corresponding to the full life cycle signal data of the rotating machinery.

[0073] Step 442 , using the mean square error between the remaining life prediction value and the corresponding true life label value as an optimization function to optimize the parameters in the initial prediction model.

[0074] Step 443: When the number of training rounds reaches the set training round value, the training is stopped to obtain an optimized initial prediction model, which is used as the lifespan prediction model.

[0075] For example, setting θ SMIFN Represents the weight and bias parameters in the prediction model (including: the weight matrices of the two attention mechanisms in the temporal information mining module, the weight matrix and bias term in the linear layer FFN; the weight matrix in the graph structure building unit in the spatial relationship modeling module, the weight matrix and bias term in the GRU network, the weight matrix in the multi-channel graph convolutional network, and the weight matrix and bias term in the linear layer of the fully connected network). Figure 3 As shown in the figure, during the model training process, the training data samples are input into the time information mining module to obtain the time features of the training data; the time features of the training data are input into the spatial relationship modeling module to obtain the spatiotemporal fusion features of the training data; the spatiotemporal fusion features of the training data and the training data samples are connected through the residual connection module to obtain the connected training data features; the connected training data features are mapped through the fully connected network to obtain the remaining life prediction value. The remaining life prediction value output by the model is and life span value y * The mean square error between the two is used as the optimization function, which is expressed as Where n batch is the batch size of input data during training, To optimize the function value, the training strategy uses the Adaptive Moment Estimation (Adam) optimizer for parameter update, calculates gradients based on the back-propagation algorithm, and achieves global optimization of model parameters through an iterative training process.

[0076] In practical applications, after obtaining the multi-source monitoring signal data within the set period T, the multi-source monitoring signal data is normalized with reference to the above step 420 and a sliding window operation is used to construct a monitoring data sample. In the following embodiment, the monitoring data sample is represented as (TW) / S+1 represents the number of sliding windows, and the monitoring data samples of each sliding window are (i.e., the monitoring data sample of the w-th sliding window), represents the set of real numbers.

[0077] As an optional implementation, in order to improve the accuracy of extracting the complex relationship between multi-source monitoring signals and the dynamic time series features of the monitoring signals, this application constructs a Figure 4 The temporal information mining module based on the dual-attention framework is shown to mine the potential inter-sequence relations and intra-sequence relations of multi-source sensor signals respectively.

[0078] In step 500, when extracting the cross-sequence relationship between multi-source monitoring signal data in the monitoring data sample in the temporal information mining module, the self-attention mechanism is used for extraction. For example, considering the heterogeneity between monitoring signal sequences collected by different sensors, it is very important to extract the cross-sequence relationship between monitoring signal sequences of different sensors. Therefore, the self-attention mechanism can dynamically assign attention weights to each monitoring signal sequence, automatically identify the importance of monitoring signal sequences collected by various sensors, and thus flexibly capture the complex inter-sequence relationship between sensors. Cross-sequence relationship H inter The extraction process is as shown in formula (1) to formula (4):

[0079] Q=X w W Q ,K=X w W K ,V=X w W V (1)

[0080]

[0081] Attn out =LayerNorm(Attn+X w )(3)

[0082] H inter =LayerNorm(FFN(Attn out )+Attn out ) (4)

[0083] Where W Q 、WK and W V represents the trainable weight matrix, is the scaling factor, FFN(·) represents the linear transformation layer, Q, K, V represent the Query, Key, and Value vectors, Attn represents the attention score, and Attn out represents the output of the attention mechanism, and LayerNorm(·) represents the layer normalization operation.

[0084] In a multi-source sensor system, the monitoring signal sequences collected by each sensor generally contain a variety of time-varying patterns, showing dynamic characteristics such as periodic changes, trend fluctuations and local anomalies. Therefore, it is of great significance to effectively extract time-varying patterns in multi-source monitoring signal data. An adaptive bidirectional autoregressive attention mechanism is proposed in the time information mining module. This mechanism can extract slow-changing trends in the steady degradation stage and fast-changing trends in the accelerated degradation stage, so that the time information mining module can accurately capture the operating state changes of rotating machinery across time windows. Based on this, in step 500, in the time information mining module, the process of extracting the intra-sequence relationship of each monitoring signal sequence in the monitoring data sample includes:

[0085] Step 511, using fast Fourier transform to obtain the time scale of the monitoring data sample based on the monitoring data sample. For example, using fast Fourier transform to dynamically analyze the main frequency components of each monitoring signal sequence in the monitoring data sample to achieve adaptive time segmentation. Considering the differences in sampling frequencies of different sensors, when performing time segmentation, the time scale s is determined. k (i.e., time segment length) as shown in formulas (5) to (7):

[0086]

[0087] f m =max(F)(6)

[0088]

[0089] Where Avg(·) represents the average operation, FFT(·) and Amp(·) represent the fast Fourier transform and amplitude value calculation respectively. represents the jth monitoring signal sequence of the wth sliding window, F represents the spectrum amplitude after Fourier transform, f m Indicates the number of divided time segments.

[0090] Step 512: A bidirectional Token sequence construction method is used to obtain a Token sequence of the monitoring data sample based on the time scale of the monitoring data sample. For example, based on the time scale s k, each monitoring signal sequence sample can be divided into multiple separate time segments. Then, a Token sequence containing both local features and global information is constructed through a cyclic translation operation, as shown in formula (8):

[0091]

[0092] Where, Token j [d,:] represents the dth token sequence of the jth monitoring signal sequence, and concat(·) represents the concatenation of vectors.

[0093] However, the sliding window size W is related to the time scale s i There is often no integer multiple relationship between them. The remaining time segments show different temporal pattern characteristics due to their inconsistent length with other time segments, which may introduce information bias or interference when calculating the attention score. Traditional cropping or padding methods will lead to information loss and dimensionality change, while the use of sliding window technology will cause segment overlap, greatly increasing the computational burden. To address these problems, this application proposes a bidirectional Token sequence construction method that segments the signal from both the forward and reverse directions, as shown in formula (9):

[0094]

[0095] Where, represents the dth forward Token sequence of the jth monitoring signal sequence, Represents the dth reverse token sequence of the jth monitoring signal sequence.

[0096] In step 513, the intra-sequence relationship of each monitoring signal data in the monitoring data sample is obtained based on the monitoring data sample and the Token sequence of the monitoring data sample. For example, the above Token sequence will be used as the key and value in the autoregressive attention mechanism, and the original sequence will be used as the query. The adaptive bidirectional autoregressive attention mechanism is shown in formulas (10) to (14):

[0097]

[0098] α=concat(α1[:k z ],mean(α o [k z :-k z ]),α2[-k z :])(12)

[0099]

[0100] H intra =LayerNorm(FFN(αout )+α out )(14)

[0101] In the formula, o=1 represents forward direction, o=2 represents reverse direction, Q o , K o 、V o Represents the Query, Key, and Value vectors of the oth direction in the proposed attention mechanism; represents the corresponding weight matrix; represents the Token sequence constructed in the oth direction of the jth monitoring sequence, α o represents the attention score in the oth direction, softmax(·) represents the softmax function, k z Indicates the interception position, mean(α o [k z :-k z ]) indicates the z To the countdown k z The attention scores of α are averaged; out represents the output of the proposed attention mechanism.

[0102] After obtaining the cross-sequence relationship and the intra-sequence relationship, the cross-sequence relationship H inter and the intra-sequence relation H intra Aggregation is performed to obtain the time characteristics of the monitoring data. For example, using the formula H = H inter +H intra , get the monitoring data time feature H, H=[H1,H2,…,H N ] T After processing the monitoring data samples of each sliding window using the time information mining module, the monitoring data time feature H∈H is obtained. ((T-W) / S+1)×N×W .

[0103] As an optional implementation, in order to improve the complex spatial relationship between the monitoring signal sequences collected by each sensor in the multi-source sensor system, the spatial relationship modeling module in step 500 includes a graph structure construction unit and a multi-channel graph convolutional network, such as Figure 5 As shown in FIG. A graph structure construction unit is used to obtain a first adjacency matrix, a second adjacency matrix, and a third adjacency matrix based on the temporal features of the monitoring data. A multi-channel graph convolutional network is used to obtain spatiotemporal fusion features of the monitoring data based on the first adjacency matrix, the second adjacency matrix, and the third adjacency matrix.

[0104] In the graph structure construction unit, the process of obtaining the first adjacency matrix, the second adjacency matrix, and the third adjacency matrix based on the time characteristics of the monitoring data includes:

[0105] Step 521, determine the cosine similarity between each two monitoring signal sequences in the time feature of the monitoring data, and generate a first adjacency matrix based on the obtained cosine similarity value. For example, similarity measurement is a common method for characterizing graph structure. In this embodiment, cosine similarity is used as an example for explanation. Cosine similarity measures the similarity of two sensor signals (i.e., monitoring signal sequences) by calculating the cosine angle between them. When the cosine similarity value is high, it indicates that the change trends of the two signals are closer, and the corresponding sensors will form a stronger topological connection in the feature graph. According to the monitoring data time feature {H1, H2, ..., H N}∈H, the adjacency matrix based on cosine similarity is shown in formula (15) to formula (16):

[0106]

[0107] Where, cos(H i ,H j ) represents the cosine value between the time feature of the monitoring data corresponding to the i-th monitoring signal sequence and the time feature of the monitoring data corresponding to the j-th monitoring signal sequence; Indicates whether there is an edge connection between the i-th sensor and the j-th sensor based on cosine similarity. Indicates that there is an edge connection. Indicates that there is no edge connection. Get the first adjacency matrix It can effectively characterize the local static spatial correlation characteristics between sensors.

[0108] In step 522, a second adjacency matrix is ​​generated based on the sensor type of the corresponding monitoring signal sequence in the time characteristics of the monitoring data. For example, in a multi-source sensor system, the sensor type can be used as domain prior knowledge and classified according to its inherent physical properties. Sensors of the same type often have stronger intrinsic correlation characteristics. This prior knowledge can effectively guide the initialization process of the second adjacency matrix, thereby improving the model's ability to represent the spatial relationship of multi-source sensors. The adjacency matrix based on sensor type is shown in formula (17):

[0109]

[0110] Where, Indicates whether there is an edge connection between the i-th sensor and the j-th sensor according to the sensor type. Indicates that there is an edge connection. Indicates that there is no edge connection; type(·) indicates the query type.

[0111] Considering that the above prior knowledge can only determine whether there is an edge connection between sensors, a self-gating mechanism is adopted to enhance the adaptive representation ability of this part and realize the dynamic learning of edge weights, as shown in formula (18):

[0112]

[0113] Where, represents the adjacency matrix considering edge weights, σ 1 is the sigmoid function, W 1 Represents the learnable weight parameter. Based on formula (18), the second adjacency matrix is ​​obtained The dependence on data volume is significantly reduced, which enables a more accurate understanding of the essential characteristics of sensor data and effectively captures the heterogeneous relationships between sensors.

[0114] Step 523, based on the time information of the monitoring signal sequence in the time characteristics of the monitoring data, a third adjacency matrix is ​​generated. For example, in actual industrial scenarios, the spatial relationship between multi-source sensor signals (i.e., monitoring signal sequences) has significant time-varying characteristics. Specifically, fluctuations in environmental parameters or changes in equipment operating conditions will cause significant changes in signal correlation. Traditional static similarity measurement methods and predefined graph methods are difficult to effectively characterize such complex spatiotemporal evolution laws. Therefore, a dynamic graph structure based on time information is proposed, and the construction process is shown in Formulas (19) to (21):

[0115] H en =σ 2 (HW 2 ) (19)

[0116]

[0117]

[0118] Where H en is the enhanced feature; 2 Represents the LeakyReLU activation function; Mask represents the operation of setting the value of the diagonal position to negative infinity to prevent the sensor from self-connecting and reduce the impact on the connection relationship with other sensors; the Softmax function is used to limit the output of the weight coefficient to the range of [0,1]; I N is the unit matrix; δ is the weight threshold. When it is less than the threshold, there is no edge connection between the i-th sensor node and the j-th sensor node. Indicates whether there is an edge connection between the i-th sensor and the j-th sensor according to the sensor type. Indicates that there is an edge connection. Indicates that there is no edge connection. Based on formula (21), the adjacency matrix A3 is obtained.

[0119] To further define the weight of the edge, a dynamic weight update mechanism based on a two-layer GRU architecture is constructed in the spatial relationship modeling module. Figure 5 As shown, the first layer of GRU network (ie GRU1) obtains temporal dynamic features along the time dimension Then the temporal dynamic feature matrix is ​​transposed and input into the second layer GRU network (i.e. GRU2), and the feature space is projected along the sensor dimension (i.e. dimension N) to mine the spatiotemporal correlation pattern across sensors. The final output representation is The dynamic process is shown in formulas (22) to (27):

[0120]

[0121] Among them, r b,t and r b,i There are two layers of reset doors; b,t and z b,i Represent the update gates of the two layers respectively; W h,1 and W h,2 are the weight matrices of the two layers respectively; b h,1 and b h,2 are the bias matrices of the two-layer GRU network; H i represents the time characteristics of the monitoring data corresponding to the input i-th monitoring signal sequence; h t-1 represents the previous hidden state; represents the candidate hidden state at time t, h b,t represents the hidden state at time t, represents the candidate hidden state of the i-th sensor sequence, h b,i represents the hidden state of the i-th sensor sequence.

[0122] The adjacency matrix A3 and the output are represented as Perform dot product processing to obtain the third adjacency matrix.

[0123] In the multi-channel graph convolutional network, the spatiotemporal fusion features of monitoring data are obtained based on the first adjacency matrix, the second adjacency matrix and the third adjacency matrix. For example, Figure 5 As shown in , the first adjacency matrix, the second adjacency matrix, and the third adjacency matrix are respectively used as three independent channel inputs of the multi-channel graph convolutional network. The spatial domain feature mining and channel feature aggregation of each channel are realized through the graph convolution with shared parameter weights. The calculation process is shown in formulas (28) to (31):

[0124]

[0125] Where A1 represents the adjacency matrix constructed based on cosine similarity, Represents the graph features of the similarity graph, A 2 represents the adjacency matrix constructed based on sensor type, The graph features representing the sensor type graph, A 3 represents the adjacency matrix constructed based on time information, represents the graph features of the time information graph, f represents the graph convolution operation, stack(·) represents the vector stacking operation, BN(·) represents the batch normalization operation, H G Represents the spatiotemporal fusion characteristics of monitoring data.

[0126] As an optional implementation, in order to alleviate the gradient vanishing problem in deep neural networks, in step 500, the monitoring data samples and the monitoring data spatiotemporal fusion features are residually connected in the residual module to obtain the connected monitoring data H out , H out =H G +X.

[0127] In an exemplary embodiment, in combination with the steps in the above embodiments, a rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network is used to test a CMAPSS data set as an example for description.

[0128] The CMAPSS dataset is one of the most influential research datasets in the field of remaining useful life prediction of turbofan engines. Figure 6 The dataset consists of four subsets, FD001-FD004, generated under different combinations of six operating conditions and two failure modes. Each subset is further divided into a training set and a test set with actual RUL labels. Details of the four subsets are shown in Table 1.

[0129] Table 1 Detailed information of the four subsets of the CMAPSS dataset

[0130] Dataset FD001 FD002 FD003 FD004 Number of training engines 100 260 100 249 Number of test engines 100 259 100 248 Number of working conditions 1 6 1 6 Number of failure mode types 1 1 2 2

[0131] The data columns include unit ID, timestamp (in cycles), three operation settings, and monitoring data of 21 sensors. Among the 21 sensors, the output values ​​of sensors 1, 5, 6, 10, 16, 18, and 19 remain constant throughout the run-to-failure experiment and cannot provide effective information for RUL prediction. Therefore, in this embodiment, only the monitoring data of the remaining 14 sensors are selected for RUL estimation. The prior map constructed based on sensor type is shown in Figure 2. Figure 7 As shown in the figure, nodes marked with the same color indicate that they have the same sensor type.

[0132] Combining the steps in the above embodiment, the RUL prediction results of EU34 in FD001, EU07 in FD002, EU94 in FD003 and EU135 in FD004 are obtained, respectively. Figure 8 (a) Figure 8 (b) Figure 8 (c) and Figure 8 (d) As shown. The test results show that no matter what degradation stage the engine is in, the predicted RUL curve maintains good consistency with the actual RUL curve, and the prediction error is within an acceptable range. Taking EU34 in FD001 as an example, although there is a certain deviation in the healthy stage, its prediction results in the degradation stage are highly consistent with the true value. This difference may be attributed to the shorter duration of the healthy stage and the difference in sensitivity of different sensors to engine degradation. The test results of EU94 in FD003 and EU135 in FD004 further verify that the solution of this application can accurately distinguish between healthy and degraded states for engines with a longer healthy stage.

[0133] For actual maintenance decisions, RUL prediction in the degradation stage is crucial. To this end, in this example, the predicted RUL value and the actual RUL value of each engine at the last checkpoint in the FD001, FD002, FD003 and FD004 test sets are compared and analyzed, respectively. Figures 9 to 12 The results show that most of the predicted RUL values ​​have only slight deviations from the true values, which verifies that the proposed life prediction model has high prediction accuracy.

[0134] Comparative test:

[0135] In the test of remaining useful life prediction, the root mean square error (RMSE) and score function (Score) are widely used quantitative evaluation indicators of model performance, which are defined as follows:

[0136] 1)RMSE:

[0137] 2) Score:

[0138] In order to fully verify the superiority of the present application, based on the above evaluation indicators, a comparative experiment was conducted between the method proposed in the above embodiment and 12 classic and cutting-edge methods. The RMSE and Score results that summarize the experimental results of each method are shown in Table 2. The methods in the table include MS-DCNN (multi-scaledeep convolutional neural network), GGCN (gated graph convolutional network), Informer (an efficient transformer-based model for LSTF), Crossformer (a Transformer-based model utilizing cross-dimension dependency for MTS forecasting), IMDSSN (integrated multi-head dualsparse self-attention) network), LOGO (Local-Global correlation fusion-basedframework), DVGTformer (dual-view graph Transformer), MSG-NET (Learning Multi-Scale Inter-Series Correlations for Multivariate Time Series Forecasting), TATFA-Transformer (Trend-augmented and temporal-featured Transformer), FC-STGNN (Fully-Connected Spatial-Temporal Graph Neural Network), ATRCN (Attention-augmentedrecalibrated and Compensatory network), SGIAEC (Structural-Guided Interaction and Attention-Enhancing Compensation Network), and the proposed SMIFN (lifespan prediction method provided by this application). Each experiment was repeated five times, and the final result is the average. The experimental results show that the prediction methods proposed in the above examples show superior prediction performance in most cases.

[0139] Table 2 RMSE and Score results of each method

[0140]

[0141] In an exemplary embodiment, in combination with the steps in the above embodiments, the performance of a rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network is evaluated using SCADA data from an onshore wind farm. Supervisory control and data acquisition (SCADA) systems are widely used for operating data collection and status monitoring of wind turbines. This embodiment uses SCADA data from an onshore wind farm to perform a performance evaluation of the method proposed in the above embodiments. The wind farm consists of 26 2.0MW semi-direct drive wind turbines, whose gearboxes use a one-stage planetary gear and two-stage helical gear transmission system with a transmission ratio of 117.66. Figure 13 The installation locations of some sensors are shown. The SCADA system collects data at a sampling interval of 1 minute. The sensor measurement information is shown in Table 3. According to the fault log records, wind turbines No. 7, No. 15, and No. 18 have been shut down due to faults and were therefore selected as experimental objects in this embodiment. The prior graph structure constructed based on sensor type is shown in Table 3. Figure 14 As shown in the figure, nodes with the same color correspond to sensors of the same type.

[0142] Table 3 Sensor measurement information

[0143] Fan number Cause of failure Operating time / day Sample size No. 7 Severe wear of high-speed shaft 468 333 No. 15 The gear ring in the gear box is severely worn 312 247 No. 18 The tooth surface of the secondary large tooth of the gearbox intermediate shaft is peeling 231 180

[0144] This embodiment uses the data of wind turbines No. 7 and No. 18 as the training set, and the data of wind turbine No. 15 as the test set. Combining the steps in the above embodiment, the remaining service life of wind turbine No. 15 is predicted, and the results are as follows: Figure 15 It is observed that the predicted RUL curve is highly consistent with the actual RUL curve, which verifies the accuracy of the proposed method.

[0145] Comparative experiment:

[0146] To further evaluate the performance of the proposed method in the above example, this example compares it with six advanced methods: Bi-LSTM (Bidirectional Long Short-Term Memory), GGCN, Transformer, Crossformer, FC-STGNN, and MSG-NET. The RMSE and Score results of these methods on wind turbine No. 15 are shown in Table 4. As can be seen from Table 4, the proposed life prediction model outperforms the other compared methods in both prediction accuracy and robustness, achieving the best RUL prediction results.

[0147] Table 4 RMSE and Score results of each method on No. 15 wind turbine

[0148]

[0149]

[0150] It can be seen from the above two embodiments that compared with classic and cutting-edge deep learning methods, the rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network proposed in this application benefits from its powerful ability to capture spatiotemporal dependencies, and is significantly superior to existing methods in the remaining service life prediction task, demonstrating superior performance.

[0151] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 16 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data related to the rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network is implemented.

[0152] Those skilled in the art will understand that Figure 16 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application and does not constitute a limitation on the computer device to which the solution of the present application is applied. A specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of the above-mentioned method embodiments when executing the computer program.

[0153] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0154] In an exemplary embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0155] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0156] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0157] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may include, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic units, data processing logic units based on quantum computing, and the like.

[0158] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0159] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for predicting the life of rotating machinery based on a multi-sensor spatiotemporal information fusion network, characterized in that: include: Obtain multi-source monitoring signal data of the rotating machinery to be predicted within a set period; The multi-source monitoring signal data includes monitoring signal sequences collected by multiple different types of sensors; performing normalization processing on the multi-source monitoring signal data; A sliding window of a set size is used to slide on the normalized multi-source monitoring signal data according to a set step size to construct a monitoring data sample; Construct a temporal information mining module, a spatial relationship modeling module, a residual connection module, and a fully connected network; Performing program optimization processing on the time information mining module, the spatial relationship modeling module, the residual connection module and the fully connected network to obtain a life prediction model; The monitoring data sample is input into the life prediction model to obtain the remaining service life value of the rotating machinery to be predicted; wherein the time information mining module is used to extract the cross-sequence relationship between the multi-source monitoring signal data in the monitoring data sample and the intra-sequence relationship of each monitoring signal sequence, and aggregate the cross-sequence relationship and the intra-sequence relationship to obtain the time feature of the monitoring data; the spatial relationship modeling module is used to obtain the spatiotemporal fusion feature of the monitoring data based on the time feature of the monitoring data; The residual module is used to perform residual connection on the monitoring data sample and the spatiotemporal fusion feature of the monitoring data to obtain the connected monitoring data; the fully connected network is used to obtain the predicted value of the remaining service life of the rotating machinery to be predicted based on the connected monitoring data.

2. The rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network according to claim 1 is characterized in that: The time information mining module, the spatial relationship modeling module, the residual module and the fully connected network are subjected to program optimization processing to obtain a life prediction model, including: Acquire multi-source monitoring signal data of the rotating machinery throughout its life cycle, and perform life annotation on the multi-source monitoring signal data of the rotating machinery throughout its life cycle to obtain training data; Normalizing the training data to obtain normalized training data; Using a sliding window of a set size, sliding on the normalized training data according to the set step size to construct a training data sample; Programming the time information mining module, the spatial relationship modeling module, the residual module, and the fully connected network to obtain an initial prediction model; The number of training rounds is set, and the initial prediction model is trained based on the training data samples. When the number of training rounds reaches the set number of training rounds, the training is stopped to obtain an optimized initial prediction model; and the optimized initial prediction model is used as the life prediction model.

3. The rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network according to claim 1 is characterized in that: A self-attention mechanism is adopted to extract cross-sequence relationships between multi-source monitoring signal data in the monitoring data samples.

4. The rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network according to claim 1 is characterized in that: In the time information mining module, the process of extracting the intra-sequence relationship of each monitoring signal sequence in the monitoring data sample includes: Obtaining a time scale of a monitoring data sample based on the monitoring data sample using a fast Fourier transform; A bidirectional Token sequence construction method is used to obtain a Token sequence of the monitoring data sample based on the time scale of the monitoring data sample; The intra-sequence relationship of each monitoring signal data in the monitoring data sample is obtained based on the monitoring data sample and the Token sequence of the monitoring data sample.

5. The rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network according to claim 1, characterized in that: The spatial relationship modeling module includes a graph structure construction unit and a multi-channel graph convolutional network; The graph structure construction unit is used to obtain a first adjacency matrix, a second adjacency matrix and a third adjacency matrix based on the time characteristics of the monitoring data; The multi-channel graph convolutional network is used to obtain the spatiotemporal fusion features of the monitoring data based on the first adjacency matrix, the second adjacency matrix and the third adjacency matrix.

6. The method for predicting the life of rotating machinery based on a multi-sensor spatiotemporal information fusion network according to claim 5, characterized in that: In the graph structure construction unit, the process of obtaining the first adjacency matrix, the second adjacency matrix, and the third adjacency matrix based on the time characteristics of the monitoring data includes: Determining the cosine similarity between each two monitoring signal sequences in the time characteristics of the monitoring data, and generating the first adjacency matrix based on the obtained cosine similarity values; generating a second adjacency matrix based on a sensor type of the monitoring signal sequence corresponding to the time feature of the monitoring data; The third adjacency matrix is ​​generated based on the time information of the monitoring signal sequence in the monitoring data time feature.

7. The method for predicting the life of rotating machinery based on a multi-sensor spatiotemporal information fusion network according to claim 2, characterized in that: Training the initial prediction model based on the training data sample, stopping the training when the number of training rounds reaches the set training round value, and using the trained initial prediction model as the life prediction model, including: Using the initial prediction model, based on the training data samples, a remaining life prediction value corresponding to the full life cycle signal data of the rotating machinery is obtained; Optimizing the parameters of the initial prediction model by using the mean square error between the remaining life prediction value and the corresponding true life label value as an optimization function; When the number of training rounds reaches the set training round value, the training is stopped to obtain an optimized initial prediction model; the optimized initial prediction model is used as the life prediction model.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the rotating machinery life prediction method based on a multi-sensor spatiotemporal information fusion network according to any one of claims 1 to 7.

9. The computer device according to claim 8, wherein: The memory is a computer-readable storage medium.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for predicting the life of rotating machinery based on a multi-sensor spatiotemporal information fusion network according to any one of claims 1 to 7 is implemented.

Images