Time-frequency-space three-level fusion industrial multi-sensor signal anomaly detection method
By employing a three-level fusion method of time, frequency, and space, combined with local-global comparative difference learning, the problem of accurate identification of abnormal feature patterns in complex industrial environments was solved, achieving high-precision and robust timestamp-level anomaly detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUNAN UNIV OF SCI & TECH
- Filing Date
- 2026-04-29
- Publication Date
- 2026-05-29
Smart Images
Figure CN122113018A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of signal processing and anomaly detection technology, and in particular to an industrial multi-sensor signal anomaly detection method based on time-frequency-space three-level fusion. Background Technology
[0002] In real-world industrial environments, the continuous rotational disturbances from equipment operation and the periodic effects of repetitive control commands often cause sensor signals to exhibit drastic fluctuations, leading to a high degree of confusion between abnormal changes and normal operating conditions, making accurate anomaly identification extremely challenging. While significant breakthroughs have been achieved in unsupervised anomaly detection, in complex industrial applications, these methods still face a core technical challenge: accurately identifying and extracting abnormal feature patterns from highly coupled and intertwined normal operating characteristics within a compact one-dimensional feature space. This bottleneck stems primarily from two factors: the inherent complexity of industrial sensor signals and the limited data space.
[0003] To address this challenge, existing solutions typically use the frequency and spatial domains as complementary perspectives to the time domain, and leverage time-frequency and spatiotemporal fusion modeling to improve detection accuracy. The core mechanism of time-frequency fusion learning lies in the fact that anomalous patterns with low separability in the time domain often exhibit stronger identifiability in the frequency domain. The core mechanism of spatiotemporal fusion modeling is that anomalies are not merely manifested as abrupt changes within a single variable, but also reveal themselves through anomalous interactions between variables. Fusing temporal evolution patterns with inter-variable dependencies in modeling can effectively improve the overall performance of anomaly detection.
[0004] Based on the fusion strategy, existing time-frequency / spatiotemporal fusion solutions are mainly divided into three categories: data-level fusion, feature-level fusion, and decision-level fusion.
[0005] (1) Data-level fusion-based detection schemes provide a feasible approach to solving the problem. This type of method relies on frequency domain and spatial domain information to supplement time domain features. By constructing time-frequency enhancement or time-space enhancement datasets, the ability of the model to identify anomalies can be improved.
[0006] (2) In addition to the data-level fusion paradigm, anomaly detection schemes based on feature-level fusion have also been widely studied. This type of method treats the time domain and frequency domain, and the time domain and spatial domain as complementary feature views, and generally adopts a dual-branch network architecture to independently learn and extract features from each view. Subsequently, the detection accuracy is improved through multi-view feature fusion strategies. Compared with data-level fusion methods, this paradigm does not require explicit data augmentation scheme design, effectively avoiding the problem of incompatibility between data augmentation and decision-making mechanisms. However, this type of method also has performance bottlenecks. Its final detection effect is highly dependent on the effectiveness of the multi-view feature fusion process. An unreasonable fusion strategy can easily lead to feature redundancy or loss of key information.
[0007] (3) Anomaly detection schemes based on decision-level fusion have also become an important research branch. Such methods also adopt a dual-branch network architecture, independently calculate anomaly scores from the perspectives of time domain, frequency domain or spatial domain, and then integrate the anomaly scores of multiple branches through decision fusion rules to output the final detection decision. The detection performance of such methods also heavily depends on the effectiveness of the decision fusion strategy. Improper strategy design will directly reduce the accuracy and reliability of the final decision.
[0008] In summary, data-level fusion is simple but lacks flexibility, while feature-level and decision-level fusion, although more flexible, suffers from detection performance highly dependent on the design quality of the fusion mechanism. Furthermore, while time-frequency and spatiotemporal fusion have been widely applied in existing research, research on time-frequency-spatial fusion learning that integrates the time, frequency, and spatial domains remains scarce. Simultaneously, current mainstream cross-domain fusion schemes generally suffer from superficial fusion mechanisms, making it difficult to deeply explore the intrinsic connections and interaction characteristics between different feature domains, significantly limiting the model's ability to capture deep cross-domain interaction information.
[0009] The contrast-based anomaly decision mechanism, primarily aimed at fine-grained timestamp-level temporal anomaly detection tasks, aims to overcome the limitations of traditional fitting-based difference decision strategies and their strong dependence on precise alignment between predicted or reconstructed values and true values. While the contrast-based difference decision mechanism has demonstrated excellent applicability and effectiveness in complex industrial signal detection, its detection performance is highly dependent on the quality of the contrast perspective construction. Current mainstream block-based contrast perspective construction methods, although ensuring basic detection stability, suffer from serious information leakage risks. This is because the local and global perspectives corresponding to a single timestamp often incorporate historical and future temporal information. When the detection timestamp arrives in real time, future data cannot be obtained in advance, making such methods unsuitable for real-time anomaly detection scenarios. Summary of the Invention
[0010] To address the shortcomings of existing technologies, this invention provides a time-frequency-space three-level fusion method for detecting anomalies in industrial multi-sensor signals. Through time-frequency-space fusion of the entire data-feature-decision process and local-global comparative difference learning, it achieves high-precision and highly robust timestamp-level anomaly detection.
[0011] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:
[0012] A time-frequency-space three-level fusion method for detecting anomalies in industrial multi-sensor signals includes the following steps:
[0013] Step S1: Acquire multi-dimensional time-series signals collected by multiple sensors during the industrial process;
[0014] Step S2: For each timestamp, based on its historical time domain context, the frequency domain amplitude spectrum obtained by Fourier transform, and the spatial correlation pattern between multiple variables, a local pattern fourth-order tensor and a global pattern fourth-order tensor are constructed respectively to achieve data-level time-frequency-space fusion.
[0015] Step S3: Perform orthogonal tensor decomposition on the local mode fourth-order tensor and the global mode fourth-order tensor to obtain multiple orthogonal two-dimensional matrices. Use a bidirectional large kernel convolutional network to extract time-domain features, frequency-domain features and spatial features from the orthogonal two-dimensional matrices. Then, concatenate and fuse the extracted features to achieve feature-level time-frequency-space fusion.
[0016] Step S4: Based on the fused spatiotemporal features, calculate the contrast difference between the local and global patterns of each timestamp from the perspectives of intra-variable context and inter-variable coupling, respectively, and then weight and fuse the contrast differences from the two perspectives to generate the final anomaly score for each timestamp, thereby achieving decision-level spatiotemporal fusion.
[0017] Step S5: Based on the comparison result between the final anomaly score and the preset threshold, determine whether the corresponding timestamp is an anomaly.
[0018] A further improvement to the above technical solution is as follows:
[0019] Preferably, step S2 specifically includes the following:
[0020] For each timestamp n, extract 2L consecutive historical timestamps forward from that timestamp to form a local context sequence; starting from timestamp n-2L, perform interval sampling at 2L steps, repeating 2L times to form a global context sequence; where 2L is the preset context length.
[0021] Perform Fourier transform on the local context sequence and the global context sequence to extract the amplitude spectrum of length L for each timestamp;
[0022] The timestamp dimension, sensor variable dimension, context sequence dimension, and amplitude spectrum dimension are stacked to construct the local mode fourth-order tensor and the global mode fourth-order tensor, respectively.
[0023] Preferably, the orthogonal tensor decomposition in step S3 adopts Tucker decomposition, which decomposes the fourth-order tensor into three mutually orthogonal two-dimensional subspace matrices along the time dimension: time-context, time-space, and time-spectrum.
[0024] Preferably, the bidirectional large-kernel convolutional network in step S3 includes:
[0025] The temporal dynamic learning branch employs a two-layer temporal convolutional network, utilizing large-size one-dimensional convolutional kernels to capture the temporal evolution features of local and global patterns along the temporal dimension.
[0026] The pattern association learning branch uses full-scale and half-scale large-size one-dimensional convolutional kernels to capture the dependencies between feature values along the pattern dimension.
[0027] The bidirectional feature fusion module is used to fuse the features output by the time-dynamic learning branch and the pattern association learning branch.
[0028] Preferably, in the temporal dynamic learning branch, the first layer of the two-layer temporal convolutional network uses a large-size one-dimensional convolutional kernel that matches the size of the context window to perform causal temporal convolution, and the second layer uses another large-size one-dimensional convolutional kernel to expand the temporal receptive field; in the pattern association learning branch, feature extraction is first performed using a full-scale large-kernel one-dimensional convolutional kernel with a length equal to the number of sensor variables M, and then further learning is performed using a half-scale large-kernel one-dimensional convolutional kernel with a length of M / 2.
[0029] Preferably, in step S4, the comparison difference calculation method of the intra-variable context perspective is as follows: the intra-variable context is reconstructed by using the first multilayer perceptron to reconstruct the local time-frequency-space fusion features and the global time-frequency-space fusion features respectively, and the difference between the reconstruction results is calculated as the first comparison difference.
[0030] Preferably, in step S4, the method for calculating the comparison difference of the coupling perspective between variables is as follows: the coupling relationship between variables is reconstructed by using a second multilayer perceptron to reconstruct the local time-frequency-space fusion features and the global time-frequency-space fusion features respectively, and the difference between the reconstruction results is calculated as the second comparison difference.
[0031] Preferably, the weighted fusion in step S4 uses a balancing hyperparameter β to weight and sum the contrast differences between the two perspectives, and the value range of the hyperparameter β is [0,1].
[0032] The time-frequency-space three-level fusion industrial multi-element sensor signal anomaly detection method provided by this invention has the following advantages compared with the prior art:
[0033] (1) The time-frequency-space three-level fusion industrial multi-sensor signal anomaly detection method of the present invention is the first to integrate time-frequency-space three-level full-process fusion. For the first time, time, frequency and spatial information are integrated into the three levels of data construction, feature extraction and decision output. It effectively solves the problem of insufficient characterization ability of single or dual-domain fusion for complex industrial signals and significantly improves the accuracy of anomaly detection.
[0034] (2) The time-frequency-space three-level fusion industrial multi-sensor signal anomaly detection method of the present invention is designed with robust local-global comparison difference decision, abandons the traditional fitting difference criterion that relies on precise alignment, and adopts local-global pattern comparison difference based on historical data without future information leakage as anomaly score. It has natural immunity to irregular fluctuations of industrial signals and greatly reduces false alarm and false detection rates.
[0035] (3) The time-frequency-space three-level fusion industrial multi-sensor signal anomaly detection method of the present invention reduces cross-domain redundancy and computational complexity through orthogonal tensor decomposition; the bidirectional large kernel convolutional network can simultaneously capture long-range time dependence and inter-pattern correlation, taking into account both global vision and local details, and achieving reasonable inference speed and low resource consumption while maintaining high accuracy.
[0036] (4) The time-frequency-space three-level fusion industrial multi-sensor signal anomaly detection method of the present invention considers the anomaly mode from two perspectives: the internal temporal context of variables and the coupling relationship between variables. Through adaptive weighted fusion, the stability and robustness of decision-making are further improved, and it performs well in both single anomaly and compound anomaly scenarios. Attached Figure Description
[0037] Figure 1 This is a schematic diagram of the detection method in this invention, where (a) is the construction of a local-global fourth-order tensor, (b) is the feature extraction based on bidirectional large kernel convolution, (c) is the learning of dual-view local-global contrast bias, and (d) is the anomaly scoring based on contrast bias.
[0038] Figure 2 The diagram shows the local-global fourth-order tensor construction of this invention, where (a) is the local and global context pattern extraction based on traditional block division, (b) is the local and global context and spectral pattern extraction of the detection method of this invention, (c) is the local fourth-order tensor, and (d) is the global fourth-order tensor.
[0039] Figure 3 This is a schematic diagram of the bidirectional large kernel convolution structure of the present invention, wherein (a) represents the three core designs, (b) represents the time dynamic learning for timestamp n, and (c) represents the pattern association learning for timestamp n.
[0040] Figure 4 Visualize the anomaly detection at time points 76700-78000 for the 13th variable on the PSM dataset in the experimental verification.
[0041] Figure 5 To verify the detection stability under different random seeds in the experiment, (a) shows the stability results under the HAI dataset, and (b) shows the stability results under the SMAP dataset. Detailed Implementation
[0042] The following provides a detailed description of specific embodiments of the present invention. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.
[0043] The present invention provides a time-frequency-space three-level fusion industrial multi-element sensor signal anomaly detection method (TFSAD), comprising the following steps:
[0044] Step S1: Acquire multi-dimensional time-series signals collected by multiple sensors during the industrial process;
[0045] An industrial production system contains M sensors, which continuously collect measurement data at N sampling times. The resulting data constitutes a multivariate time series. Each row corresponds to a sensor channel, and each column corresponds to a timestamp. Indicates size is A space of real matrix numbers.
[0046] From a structural perspective, multivariate time series naturally exhibit two complementary dimensions: the time dimension and the variable dimension. In the time dimension, ,in Table 1 n A vector of observations collected by all sensors at each timestamp. In terms of variables, ,in Let m represent the observation vector of the m-th sensor at all timestamps.
[0047] Step S2: At the data level, for each timestamp, based on its historical time domain context, the frequency domain amplitude spectrum obtained by Fourier transform, and the spatial correlation pattern between multiple variables, a local pattern fourth-order tensor and a global pattern fourth-order tensor are constructed respectively, forming a set of comparative fourth-order tensors to achieve data-level time-frequency-space fusion.
[0048] Anomaly decision-making methods based on contrastive differences require each timestamp to have two corresponding sets of feature representations. In industrial sensor signal scenarios, abnormal states often persist across consecutive timestamp sequences, exhibiting characteristic patterns distinct from normal equipment operation. Based on this characteristic, this invention constructs two complementary representation patterns for a single timestamp: one is a local pattern extracted from adjacent timestamps, and the other is a global pattern constructed based on long-term time-interval timestamps. This design approach is intuitive and effective. For abnormal timestamps, there is a significant deviation between their local and global patterns, resulting in substantial feature differences. Conversely, for normal timestamps, constrained by the inherent repetitive nature of industrial production processes, the local and global patterns maintain a high degree of consistency, with the difference between them remaining at a low level.
[0049] Block partitioning is the mainstream method for extracting local and global patterns in the current time series field. Its core mechanism is to reshape a one-dimensional time series into a two-dimensional matrix with fixed block length. For a given timestamp, adjacent elements in the same row represent its local pattern, while adjacent elements in the same column represent its global pattern. The specific construction form is as follows: Figure 2 As shown in (a) above. Although the block-based method has the advantages of simplicity and high computational efficiency, it suffers from a core flaw: future information leakage. The extracted local and global patterns often contain future timestamp data that cannot be obtained during the inference stage, making it difficult to adapt to real-time anomaly detection scenarios. To overcome the above technical bottlenecks, this invention proposes an improved block-based local-global pattern extraction strategy, achieving performance upgrades through two key optimizations, as shown in Figure 1. Figure 2 As shown in (b) above. First, the local and global patterns for each timestamp are constructed solely based on historical time-series data, thus eliminating the possibility of future information leakage at its source. Second, it breaks through the limitations of a single time context by simultaneously capturing complementary temporal, frequency-domain, and spatial coupling feature information for each timestamp. These multi-dimensional complementary information are then fused along the time dimension to construct local and global fourth-order tensors, further enhancing the discriminability of the pattern representation.
[0050] The specific construction process of the fourth-order tensor for local-global comparison is as follows:
[0051] S2-1, Time Dimension: Multivariate Time Series The N timestamps are defined as the first dimension of the local and global fourth-order tensors.
[0052] S2-2, Spatial Dimension: The M variables in X are defined as the second dimension of the fourth-order tensor, representing spatial (between variables) information.
[0053] S2-3, Context Dimension: For each timestamp n in the time dimension, construct local and global context sequences to form the third dimension of the fourth-order tensor. Specifically, for each variable, construct the local context by using n as the endpoint and extracting 2L consecutive historical timestamps. Simultaneously, using n-2L as the endpoint, sample a timestamp at intervals of 2L steps forward, repeating 2L times to construct a global context of alternating lengths. Here, 2L is the context length, a predefined hyperparameter used to specify the context length. By default, the local and global contexts have the same length to ensure fairness.
[0054] S2-4, Frequency Domain Dimension: For the local and global context sequences of each timestamp, a Fast Fourier Transform (FFT) is applied to transform the context to the frequency domain. Then, the corresponding amplitude spectra are extracted to generate local and global amplitude vectors of length L, which serve as the fourth order of the fourth-order tensor.
[0055] S2-5, Tensor Formation: Following the steps above, the multivariate time series X is transformed into two fourth-order tensors: a local tensor... and a global tensor ,like Figure 2 As shown in (c) and (d), T, S, C, and A represent the temporal, spatial, contextual, and magnitude dimensions, respectively.
[0056] Step S3: At the feature level, orthogonal tensor decomposition is performed on the local mode fourth-order tensor and the global mode fourth-order tensor to obtain multiple orthogonal two-dimensional matrices. Then, a bidirectional large kernel convolutional network is used to extract time-domain features, frequency-domain features and spatial features from the orthogonal two-dimensional matrices. The extracted features are then spliced and fused to achieve feature-level time-frequency-space fusion.
[0057] One direct approach to extracting time-frequency-space fusion features is to directly apply a fourth-order kernel to a fourth-order tensor. However, this strategy makes feature extraction difficult and incurs high computational costs. Furthermore, the fourth-order tensor contains a large amount of redundant information, leading to redundancy in the feature representation. To address these issues, this step employs optimization design in two aspects:
[0058] On the one hand, a divide-and-conquer strategy is adopted. The fourth-order tensor is decomposed along the time dimension into three orthogonal second-order matrices, representing time, frequency domain, and spatial information, respectively. Through orthogonal decomposition, the high-order redundant tensor is transformed into a low-order orthogonal representation, thereby reducing computational complexity and simplifying subsequent feature fusion.
[0059] On the other hand, a customized bidirectional large-kernel convolution was designed. To enhance the ability to extract features from each orthogonal matrix, a customized convolution was designed. Specifically, a large kernel is used for feature extraction of the target features, while a bidirectional learning mechanism captures temporal dynamics and pattern associations. Furthermore, a shared local-global convolution kernel is employed to facilitate effective contrastive learning between local and global representations.
[0060] To this end, this invention designs a feature extractor based on bidirectional large kernel convolution to capture spatiotemporal fusion features from global and local fourth-order tensors. The specific steps are as follows:
[0061] S3-1, Orthogonal Tensor Decomposition: To reduce the learning difficulty and computational cost, this step applies orthogonal Tucker decomposition to local and global comparison fourth-order tensors.
[0062] The tensor is decomposed along the time dimension into three mutually orthogonal subspaces: time-context, time-space, and time-spectrum, which respectively capture the spatial private features, context private features, and spectrum private features under each timestamp.
[0063] The specific process is as follows:
[0064] S3-1-1, Fourth-order expansion of a tensor. To obtain the core tensor, we need to combine the local and global fourth-order tensors into a local tensor. and a global tensor Expanding along each order, we obtain four expansion matrices:
[0065] (1)
[0066] Where N is the number of timestamps, M is the number of sensors, 2L is the context length, and L is the dimension of the amplitude vector. and These are the matrices expanded along the time modulus in local and global modes, respectively. and These are the matrices expanded along the spatial module in local and global modes, respectively. and These are the matrices expanded along the context module in local and global modes, respectively. and These are the matrices expanded along the magnitude modulus in local and global modes, respectively.
[0067] S3-1-2, orthogonal basis generation. Perform truncated singular value decomposition on each expanded matrix:
[0068] (2)
[0069] in, and These are the time-modal orthogonal bases for local and global modes, respectively. and These are the singular value matrices obtained from the time modulo in local and global modes, respectively. and These are the transposes of the right singular vector matrices corresponding to the time modulo in local and global modes, respectively. and These are spatial module orthogonal bases for local and global modes, respectively. and These are the singular value matrices obtained from the spatial moduli in local and global modes, respectively. and These are the transposes of the right singular vector matrices corresponding to the spatial moduli in local and global modes, respectively. and These are the context module orthogonal bases for local and global modes, respectively. and These are the singular value matrices obtained from the context modulo in local and global modes, respectively. and These are the transposes of the right singular vector matrices corresponding to the context modulo in local and global modes, respectively. and These are the amplitude magnitude orthogonal bases for local and global modes, respectively. and These are the singular value matrices obtained from the magnitude moduli in local and global modes, respectively. and These are the transposes of the right singular vector matrices corresponding to the magnitude magnitudes in local and global modes, respectively. This is for the transpose operation.
[0070] By preserving the singular vectors of the principal components with a fixed rank, four orthonormal bases are obtained:
[0071] (3)
[0072] in, , , , These are the ranks retained for the time modulus, spatial modulus, context modulus, and magnitude modulus, respectively. To simplify calculations, [the following is omitted as the original text is incomplete and requires further context]. All are set to 1, and the following conditions are met:
[0073] (4)
[0074] Where 1 represents the identity matrix, It is the transpose of the time-mode orthogonal basis in the local mode. For the transpose of the time-modular orthogonal basis in global mode, For the transpose of the spatial modulus orthogonal basis under local modes, For the transpose of the spatial modular orthogonal basis in global mode, For the transpose of the context module orthogonal basis in the local mode, For the transpose of the context module orthogonal basis in global mode, For the transpose of the orthogonal basis of magnitude modulus in the local mode, This is the transpose of the magnitude modulus orthogonal basis in the global mode. These orthogonal bases form four mutually orthogonal subspaces, thus achieving a low-redundancy, interpretable tensor decomposition.
[0075] S3-1-3, construct three time-oriented orthogonal matrices. Based on the obtained orthogonal basis, construct three mutually orthogonal time-oriented subspaces: time-context, time-space, and time-spectrum, which are used to characterize the unique contextual features, spatial coupling relationships, and spectral correlation properties of each timestamp, respectively.
[0076] (5)
[0077] in, This represents the context-private feature for each timestamp in local mode. This represents the context-private feature for each timestamp in global mode. For local fourth-order tensors, It is a global fourth-order tensor. This represents the spatial coupling feature of each timestamp in the local mode. The spatial coupling feature of each timestamp in global mode. For the frequency domain spectral features of each timestamp in local mode, This represents the frequency domain spectral features of each timestamp in global mode. For tensor and matrix multiplication on the second modulus, This is a tensor-matrix multiplication over the third modulus. This is a tensor-matrix multiplication on the fourth modulus. , , .
[0078] To construct a richer enhanced representation, these three orthogonal matrices are further concatenated, as follows:
[0079] (6)
[0080] in, Enhanced features in local patterns For enhanced features in global mode, For splicing operations, , .
[0081] After decomposition, the local and global fourth-order tensors are each transformed into four orthogonal second-order matrices. and .
[0082] S3-2, Bidirectional Large Kernel Convolution Feature Extraction: Each generated local and global orthogonal matrix contains two types of complementary features: temporal dynamics and pattern correlation. Temporal dynamics represents the evolution of patterns over time, while pattern correlation reflects the potential dependencies between feature values. To accurately capture these two core features, this step designs a bidirectional large kernel convolutional network as a feature extractor. This network incorporates three core design elements, the specific structure of which is as follows: Figure 3 As shown in (a): First, large convolutional kernels are used to model representative patterns in the data; second, the convolutional kernels achieve parameter sharing in the local and global comparison learning process to ensure the consistency of pattern learning; third, branches are designed for temporal dynamics and pattern correlation, and temporal dynamic features are captured by two layers of temporal convolutional networks (TCN) stacked along the time dimension, and pattern correlation features are mined by a dual-scale one-dimensional convolutional neural network (1D-CNN) built along the pattern dimension.
[0083] With local and global orthogonal matrices For example, the bidirectional feature extraction process is as follows:
[0084] S3-2-1, Temporal Dynamic Learning: Employs a two-layer Temporal Convolutional Network (TCN) and uses large-size one-dimensional convolutional kernels for cascaded temporal convolution.
[0085] First, for each timestamp and channels Using large-scale one-dimensional convolution kernels , Its length corresponds to the size of the context window, within the history window. Perform causal temporal convolution on top. This operation can capture latent temporal dynamic features at the timestamp level, such as... Figure 3 As shown in (b):
[0086] (7)
[0087] in, These are intermediate features after passing through the first TCN layer in local mode. These are intermediate features after passing through the first layer of TCN in global mode. This represents the spatial coupling feature of each timestamp in the local mode. The spatial coupling feature of each timestamp in global mode. The large core of the first-layer TCN, is the element index in the convolution kernel, and i is the position index in the temporal convolution window. subscript This represents the element in the nth row and mth column of the matrix.
[0088] Secondly, cross-channel interactions among M channels are introduced to enhance feature representation:
[0089] (8)
[0090] in, This represents the parameters of the first-layer TCN, and LeakyRelu() is used as the activation function. These are features obtained through cross-channel interaction after passing through the first layer of TCN in local mode. These are features obtained through cross-channel interaction after passing through the first layer of TCN in global mode.
[0091] Finally, through a second large one-dimensional convolution kernel ,and A second-layer TCN is executed to expand the temporal receptive field and capture higher-order temporal dynamic features:
[0092] (9)
[0093] in, These are intermediate features after passing through the second TCN layer in local mode. These are intermediate features after passing through the second layer of TCN in global mode. The kernel is the large kernel of the second temporal convolution. .
[0094] Finally, channel-by-channel interaction among the M channels is introduced to further optimize the temporal feature representation:
[0095] (10)
[0096] in, This represents the parameters of the second-layer TCN, and ; The temporal dynamic features are extracted from two layers of TCN in local mode. The temporal dynamic features are extracted by a two-layer TCN in global mode. For features resulting from the second layer cross-channel interaction in local mode, For features resulting from the second layer of cross-channel interaction in global mode, This is the cross-channel parameter matrix of the second-layer TCN.
[0097] After the above operations, local and global orthogonal matrices are respectively generated. Finally, the temporal features are obtained. .
[0098] S3-2-2, Pattern Association Learning: Employing a large kernel of length M (equal to the pattern dimension) with full-scale dimensions. A one-dimensional convolutional neural network for each timestamp Extract the correlation between values along the pattern dimension, and ,like Figure 3 As shown in (c). Represented as:
[0099] (11)
[0100] in, The features obtained by full-scale convolution in local mode are... The features are obtained by full-scale convolution in global mode. For the local spatial coupling feature at the nth timestamp, the jth index, For the global spatial coupling feature at the nth timestamp, the jth index, It is a full-scale mode convolution kernel, and .
[0101] Then, another half-scale convolution kernel was used. ,and Its length is M / 2, and at each timestamp Further learning mode association:
[0102] (12)
[0103] in, For the local pattern association feature value learned at the nth timestamp and the kth sliding position, For the global pattern associated feature value learned at the nth timestamp and the kth sliding position, This represents the element value at the (k+j-1)th pattern position of the nth timestamp in the local pattern. This refers to the element value at the k+j-1th pattern position of the nth timestamp in global mode. It is a half-scale mode convolution kernel. It is an index value, and .
[0104] at last, The final associated features are generated as follows:
[0105] (13)
[0106] in, These are local pattern association features obtained by a dual-scale feature extractor. These are global pattern association features obtained by a dual-scale feature extractor.
[0107] S3-2-3, Bidirectional Feature Fusion: By fusing bidirectional features, a local and global orthogonal matrix is obtained. The final feature representation:
[0108] (14)
[0109] in, For the final local time-space characteristics, For the final global time-space features, .
[0110] Through the above steps, four types of orthogonal matrices, both local and global, are obtained. and They are respectively converted into their corresponding feature representations: and ,in, For local time-context features, It is a local time-spectral feature. To enhance the splicing features locally, For global time-context features, For global time-spectral characteristics, This enhances the splicing features globally.
[0111] S3-2-4, Time-Frequency-Spatial Feature Fusion: Based on orthogonality, the local and global time-frequency-spatial fusion features can be concisely generated as follows:
[0112] (15)
[0113] in, The final time-frequency-space fusion feature in the local mode, This represents the final time-frequency-space fusion feature in global mode, and .
[0114] Step S4: At the decision level, based on the fused spatiotemporal features, the contrast difference between the local and global patterns of each timestamp is calculated from the perspectives of intra-variable context and inter-variable coupling, respectively. The contrast differences from the two perspectives are then weighted and fused to generate the final anomaly score for each timestamp, thus achieving spatiotemporal fusion at the decision level.
[0115] Currently, most mainstream anomaly detection methods employ independent decision-making mechanisms. These mechanisms generate timestamp-level decision results (anomaly scores) for individual variables, and then filter or aggregate these results to arrive at a final judgment. Such independent decision-making strategies rely solely on single temporal context information for modeling, neglecting the inherent coupling relationships between multiple variables. However, in real-world industrial multivariate monitoring systems, anomalous events are typically triggered by anomalous interactions between variables, rather than isolated signals from a single sensor. To address this deficiency, this step adopts a variable-dependent decision-making strategy based on two complementary perspectives: an intra-variable context perspective and an inter-variable coupling perspective. The intra-variable context perspective focuses on the temporal characteristics within a single variable, uncovering local-global contrast differences between timestamps; the inter-variable coupling perspective focuses on the correlation characteristics of multiple variables, capturing contrast differences caused by variable interactions. By fusing the decision outputs from these two perspectives, this step constructs a decision-level time-frequency-spatial fusion framework, thereby achieving high-precision and robust industrial temporal anomaly detection.
[0116] Specifically, it includes the following:
[0117] S4-1, Intra-variable context-driven contrastive difference learning: Based on learned local and global spatio-temporal fusion features and A two-layer multilayer perceptron network with one input neuron of 3M+9L+8, one hidden neuron of d, and one output neuron of M×2L is used for each timestamp. Restore the context within the variables:
[0118] (16)
[0119] in, The output of the context within the variables obtained by restoring the local branch at the nth time stamp. This is the context output of the variables recovered at the nth timestamp of the global branch, and , . Let be the weight matrix of the second-layer MLP, and , Let be the weight matrix of the first layer MLP, and , For the local time-frequency-space fusion feature of the nth timestamp, The bias of the first layer MLP, and , For the bias of the second layer MLP, and . Let be the global time-frequency-spatial fusion feature of the nth timestamp, and , It is also used as the input to a two-layer MLP.
[0120] S4-2, Contrastive Difference Learning Driven by Inter-Variable Coupling: Local and Global Spatiotemporal Fusion Features and A two-layer multilayer perceptron network with another input neuron of 3M+9L+8, a hidden neuron of d, and an output neuron of M is used for each timestamp. Restore the coupling relationship between variables:
[0121] (17)
[0122] in, This is the output of inter-variable coupling recovered at the nth timestamp in local mode. This is the output of variable coupling recovered at the nth timestamp in global mode, and , . Let be the weight matrix of the second-layer MLP, and . Let be the weight matrix of the first layer MLP, and . The bias of the first layer MLP, and . For the bias of the second layer MLP, and , It is used as the input to a two-layer MLP.
[0123] S4-3, Loss Function: The loss function for dual-view contrastive learning is defined as follows:
[0124] (18)
[0125] in, and Represents the actual value. For the total loss function, The loss is the loss from the perspective of the variable's internal context. The loss is due to the coupling between variables. For the context target output within variables in local mode, For the context target output within variables in local mode, This outputs the coupling between variables in local mode. This outputs the coupling between variables in local mode.
[0126] S4-4, Anomaly scoring based on contrast difference: After training, two contrast difference values are generated for each test timestamp. These two differences are fused at the decision layer through the anomaly scoring mechanism to obtain the final anomaly score.
[0127] (19)
[0128] in, Let be the final anomaly score for the t-th test timestamp. Output the value at the t-th timestamp and the j-th element for the local context. Output the value at the t-th timestamp and the j-th element to the global context. To achieve local coupling, output the value of the m-th variable at the t-th timestamp. To globally couple the output value at the t-th timestamp and the m-th variable, This represents the local-global contrast difference driven by the intra-variable context from perspective 1. This represents the local-global contrast difference driven by inter-variable coupling from perspective 2; It is a hyperparameter used to balance the two types of scores, and its value ranges from [value range missing]. The smaller it is, the more emphasis is placed on it. ,on the contrary The higher the proportion.
[0129] Step S5: Based on the comparison result between the final abnormal score and the preset threshold, determine whether the corresponding timestamp is abnormal.
[0130] abnormal scores With preset threshold Compare these to determine if any point in time is abnormal:
[0131] (20)
[0132] in this case, Indicates a point in time It was identified as abnormal, and This indicates a normal state.
[0133] Experimental verification:
[0134] As shown in Table 1, eight publicly available real-world industrial sensor multivariate signal datasets were used to validate the proposed method. These datasets cover a wide range of industrial and IoT fields, ensuring the diversity and fairness of the experimental evaluation.
[0135] Table 1. Eight real-world industrial sensor multivariate signal datasets
[0136]
[0137] The SKAB dataset provides multi-sensor signals collected from industrial testbeds. The GECCO dataset is a water quality monitoring dataset collected from IoT sensors. The PSM dataset records multi-sensor signals used for equipment health monitoring under various operating conditions. The HAI dataset is an industrial dataset containing signals from pumps, valves, pressure sensors, and other common industrial equipment. The SMD dataset records resource utilization trajectories from 28 servers over five weeks, including 38 monitoring metrics. The Genesis dataset is a comprehensive industrial dataset collected from a small-scale plant model. Additionally, the SMAP dataset provides telemetry data from NASA's Mars rovers and soil measurement data. Finally, the SWAT dataset contains 51-dimensional sensor data from a water treatment plant testbed.
[0138] To verify the effectiveness of the Time-Frequency-Spatial Three-Level Fusion Driven Detection Method (TFSAD) of this invention, nine representative state-of-the-art methods were selected as benchmark models, and their details are shown in Table 2. These benchmark models are divided into three categories: (1) Deep models using local-global contrast strategies: TFSAD and DCdetector use local-global contrast differences for anomaly detection, while the other baseline models rely on strategies based on fitting differences. TFSAD performs contrast learning at the data, feature, and decision levels, while DCdetector only performs contrast learning at the feature level. (2) Time-Frequency-Spatial Fusion Learning: TSINR, TFCL, and TFMAE focus on time-frequency fusion modeling, while STEN, GDCMAD, and GST-GL emphasize spatiotemporal fusion modeling. In contrast, TFSAD integrates time-frequency-spatial fusion learning, which can more comprehensively represent complex industrial patterns. (3) Deep architecture and lightweight architecture: COUTA and PatchAD are lightweight models designed to improve efficiency, while the other baseline models are built on deep neural network architectures.
[0139] Table 2. Benchmark Model
[0140]
[0141] To comprehensively evaluate the proposed method, six widely used metrics were employed: Accuracy (ACC), Precision, Recall, Point-Adjusted F1 (F1-PA), VUS_ROC, and VUS_PR. ACC calculates the overall proportion of correctly classified samples; Precision measures the proportion of predicted normal time points that are actually normal, while Recall measures the proportion of actually detected normal time points; F1-PA balances precision and recall and is adjusted based on point-by-point detection error; VUS_ROC extends the traditional ROC curve to three dimensions by introducing auxiliary evaluation parameters and measures its closed volume; VUS_PR calculates the volume under the precision-recall surface, highlighting performance on imbalanced datasets. All metrics range from 0 to 1, with higher values indicating better anomaly detection performance.
[0142] The TFSAD proposed in this invention is implemented in Python within the PyCharm IDE, and its source code has been publicly released on GitHub. All experiments were conducted on a Windows 11 platform equipped with an Intel® Core™ i7-10700KF CPU, 64GB AMOLED display, and an NVIDIA GeForce RTX 3090 GPU with 24GB of video memory. For fair comparison, the benchmark models listed in Table 2 were reproduced using the official Python implementation provided in the author's code repository.
[0143] (a) Accuracy Experiment
[0144] To conduct accuracy experiments, six evaluation metrics were used to compare the TFSAD of this invention with nine benchmark models on eight real-world industrial sensor multivariate signal datasets.
[0145] (1) Accuracy: Table 3 summarizes the anomaly detection results of the ten methods on eight datasets. Bold text indicates the best performance, and underlined text indicates the second-best performance. The following observations can be drawn from this:
[0146] Overall Performance: TFSAD consistently achieves the top detection performance on most datasets and evaluation metrics, followed by TFCL, TFMAE, and TSINR. For example, on the SKAB dataset, TFSAD's accuracy (ACC) is 0.9990 and its Fl score is 0.9986, surpassing GDCMAD (accuracy 0.9744, Fl score 0.9655) and TSINR (accuracy 0.9828, Fl score 0.9798), while its performance is comparable to PatchAD (accuracy 0.9979, Fl score 0.9970). These results demonstrate that TFSAD is highly effective in identifying anomalous patterns under complex industrial conditions.
[0147] Stability: TFSAD, TFMAE, and TFCL demonstrate relatively stable performance across various datasets. Notably, TFSAD ranks first on most metrics for the SKAB, GECCO, SMD, and SMAP datasets, while some baseline methods exhibit significant performance fluctuations.
[0148] Local-Global Contrast Difference: Models based on local-global contrast difference (TFSAD and DCdetector) typically achieve better detection performance than other methods. For example, on the SMAP dataset, DCdetector has a V_ROC of 0.9816 and a V_PR of 0.9560, outperforming STEN (0.9727 and 0.9468, respectively) and COUTA (0.6127 and 0.6295, respectively). These results demonstrate the effectiveness of local-global contrast difference learning in capturing anomalous patterns in unsupervised settings.
[0149] Spatiotemporal fusion learning: Compared to single time-domain or frequency-domain learning, time-frequency and spatiotemporal learning methods generally improve detection accuracy; however, they are still inferior to spatiotemporal fusion learning. On the PSM dataset, TFSAD achieves V_ROC and V_PR scores of 0.9231 and 0.9314, respectively, outperforming time-frequency fusion learning methods (such as TFCL's V_ROC of 0.9089 and V_PR of 0.9250, and TFMAE's V_ROC of 0.9048 and V_PR of 0.9172) and spatiotemporal fusion learning methods (including STEN's V_ROC of 0.9129 and V_PR of 0.9222, and GDCMAD's V_ROC of 0.6227 and V_PR of 0.6879). These results indicate that time, frequency, and space provide complementary perspectives, and their fusion modeling is particularly effective in improving anomaly detection performance.
[0150] Deep vs. Shallow Structures: Models based on deep structures generally exhibit stronger detection capabilities than models based on lightweight structures. On the HAI dataset, TFSAD achieved an ACC of 0.9949 and a Rec of 1, surpassing COUTA (ACC of 0.9914, Rec of 0.9915) and PatchAD (ACC of 0.9902, Rec of 0.9904).
[0151] Table 3 Accuracy Analysis
[0152]
[0153] (2) Visualization: Figure 4 This figure shows the anomaly detection results of ten models on the 13th variable of the PSM dataset at time points 76,700 to 78,000. The black curve represents the ground truth value, and the red curve represents the true / predicted label (peaks indicate anomalies, and valleys indicate normal states). Two main observations can be drawn from the figure:
[0154] Accuracy: TFSAD, DCdetector, and TFCL outperformed other models in detection accuracy. Specifically, their predictions were highly consistent with the true labels at 76,800 to 77,200 time points. In contrast, the remaining models produced multiple misclassifications, either missing true anomalies or generating false detections.
[0155] False positive rate: In the time range of 77,200 to 78,000, TFSAD generated significantly fewer false positives than other models, indicating higher accuracy in distinguishing between anomalous and normal patterns. Specifically, TFSAD generated 6 false positives, while DCdetector generated 7, TFCL 9, STEN 12, and TFMAE 15, with the remaining models generating more false positives.
[0156] Experimental results demonstrate that the TFSAD method of this invention achieves excellent and stable performance in anomaly detection. Specifically, the method consistently ranks among the top in six evaluation metrics across multiple datasets, outperforming several deep learning models.
[0157] (II) Stability Test
[0158] To verify the generalization stability of the proposed method, a model stability criterion was established: if the neural network black-box model can maintain excellent detection performance under different random seed conditions, then the model is considered stable. Based on the above criterion, a random seed mutation experiment was designed to evaluate the stability of the TFSAD model. The specific experimental procedure is as follows: Ten different random seeds were then generated, and the detection performance of the model under each random seed configuration was evaluated. This experiment selected the HAI and SMAP datasets for validation, and six core indicators were used to complete the quantitative evaluation. Figure 5 The visualization results for model stability testing show a positive correlation between the darkness of the color blocks and the degree of performance fluctuation; lighter color blocks indicate smaller performance fluctuations, while darker color blocks indicate larger performance fluctuations. For a single model, frequent changes in color blocks within the same row indicate higher performance variability across different metrics, while evenly distributed color blocks indicate better model performance consistency. Figure 5 As shown, three core observational conclusions can be extracted from the experimental results:
[0159] (1) Overall stability dimension: The overall stability of TFSAD, TSINR, CST-GL and PatchAD models is better than the other comparison models. The color patches of the four models consistently show a light color distribution on all four datasets. In contrast, the performance fluctuation of TFCL, GDCMAD and TFMAE models is larger, and the corresponding color patches show a darker overall tone.
[0160] (2) Stability across metrics: DCdetector, TFMA, GDCMAD, and COUTA exhibit weak stability across different evaluation metrics, as evidenced by the significant color variations within the same dataset. For example, on the SWAT dataset, GDCMAD shows a large fluctuation of 25.6 on Pre, while its fluctuation on ACC is only 1.5, indicating significant inconsistency between different metrics.
[0161] (3) Cross-dataset stability: STEN, DCdetector, and TFCL exhibit significant performance differences across different datasets. For example, STEN shows large fluctuations across all six evaluation metrics on the HAI dataset, while its performance changes significantly less on the other three datasets. In summary, TFSAD demonstrates superior stability compared to other models.
[0162] (III) Timeliness and resource consumption experiment
[0163] Under the same settings (batch size = 64, time window = 100), the timeliness and resource consumption of TFSAD were compared with nine baseline models. All experiments were conducted on the GECCO dataset to ensure the fairness of the comparison. The efficiency metrics used included training interval, inference time, inference time, model parameters, CPU / GPU utilization, RAM / VGA-RAM utilization, and floating-point operations (FLOPs). The experimental results are shown in Table 4.
[0164] (1) Timeliness: TFSAD demonstrates good execution efficiency by maintaining a compact model size and reasonable training and inference times. Overall, among all the compared methods, it ranks second in model size and is at a moderate level in training and testing speed, indicating that it achieves a good balance between efficiency and performance. Specifically, TFSAD contains only 162,000 parameters, far fewer than TSINR (3900K), TFCL (1216K), and GDCMAD (6790K). In terms of training efficiency, TFSAD requires 68.86 seconds per epoch. Although several models, such as TFCL, COUTA, TFMAE, STEN, and GDCMAD, train faster, taking 20-50 seconds per epoch, TFSAD still achieves a competitive training speed due to its stronger detection performance. During the inference phase, TFSAD completed the test in just 44.5 seconds, significantly faster than DCdetector (83.87 seconds) and STEN (110.2 seconds), demonstrating its practical advantages in real-time or near-real-time scenarios.
[0165] (2) Resource Consumption: TFSAD has low computational and memory overhead. Specifically, its GPU utilization rate remains at 12.1%, significantly lower than TFMAE (36.3%), TFCL (24.9%), and PatchAD (42.5%). In terms of memory consumption, TFSAD uses only 1.9GB of video memory, which is among the lowest levels in the comparison models and significantly lower than TFCL (10.6GB). Its memory utilization rate is 4.7GB, which is at a moderate level and lower than some larger models such as PatchAD (7.2GB).
[0166] (3) Computational cost: TFSAD requires only 5.4 GFLOPs, far less than almost all competing models. Specifically, it requires less than one-third of the FLOPs of DCdetector (17.79 GFLOPs) and far less than TSINR (128 GFLOPs).
[0167] Table 4 Timeliness and Resource Consumption Analysis
[0168]
[0169] (iv) Case Study Experiment
[0170] Experiments were conducted using a drive system fault and anomaly simulation test bench.
[0171] (1) Physical test bench: The physical test bench includes a motor, gearbox, magnetic powder brake, motor controller, edge data acquisition unit and control software. It is equipped with a torque sensor, a triaxial vibration accelerometer and a current / voltage clamp, which can simultaneously acquire multi-source electrical and mechanical signals, including current, voltage, vibration and torque.
[0172] (2) Data Collection and Anomaly Generation: Data collection and dataset construction were carried out using the existing testing platform. Multi-sensor time-series signals were collected at a sampling rate of 12.8 kHz, with a single collection duration of 30 seconds. Each collection record contained six synchronous data streams, namely x, y, and z-axis vibration signals, as well as torque, current, and voltage signals. To simulate abnormal working conditions in real industrial scenarios, artificial abnormal data was injected using control software. Specifically, every 12 seconds, a 3-second normal signal segment was replaced with tampered data. This anomaly injection operation was repeated four times in each record. Based on the above data collection and anomaly simulation process, two types of specialized anomaly datasets were constructed: one was a single anomaly dataset containing only broken tooth fault samples; the other was a mixed anomaly dataset, simultaneously covering both broken tooth faults and outer ring faults. The final constructed single dataset contained a total of 384,000 timestamps, of which 268,800 timestamps were allocated as the training set for model training, and the remaining timestamps were allocated as the test set for model performance evaluation.
[0173] (3) Experimental results: Table 5 shows the detection results on single anomaly datasets and mixed anomaly datasets.
[0174] On a single outlier dataset: TFSAD ranked first in five of the six evaluation metrics and second in the remaining one, demonstrating strong robustness and comprehensive performance advantages. In particular, TFSAD showed significant advantages in V_ROC and V_PR, reaching 0.9986 and 0.9917 respectively. In contrast, PatchAD's accuracy (ACC) was 0.975, and TSINR's precision (Pre) was 0.9831, both inferior to TFSAD's performance.
[0175] Hybrid anomaly datasets: Due to the increased complexity of hybrid and compound anomalies, detecting them is significantly more challenging. Nevertheless, TFSAD maintains excellent performance, ranking first across all five evaluation metrics for hybrid anomaly datasets. For example, TFSAD achieves an F1 score of 0.9897, while DCdetector scores 0.9846, TFCL 0.8648, and CST-GL only 0.6146.
[0176] These results clearly demonstrate the effectiveness and efficiency of the time-frequency-space three-level fusion-driven detection method of the present invention in time-point anomaly detection, which means that it is suitable for practical industrial applications.
[0177] Table 5. Anomaly detection results for single and composite anomalies.
[0178]
[0179] This invention proposes an anomaly detection method based on time-frequency-space three-level fusion, and constructs a time-frequency-space coupled local-global contrastive difference learning framework for industrial multi-sensor scenarios. Addressing the shortcomings of existing research on temporal anomaly detection in multi-domain feature fusion and mining, this invention introduces a time-frequency-space three-level fusion mechanism into the temporal anomaly detection task, providing a new approach for high-precision and robust anomaly monitoring in industrial scenarios.
[0180] (1) In view of the inherent characteristics of industrial sensor signals, this invention introduces an anomaly decision mechanism based on local-global comparison difference to replace the traditional fitting difference judgment method, so as to realize fine-grained timestamp-level anomaly detection.
[0181] (2) To ensure detection accuracy, a three-level fusion model of time, frequency and space is adopted instead of the traditional time-frequency or time-space modeling, so as to enhance the identifiability of abnormal patterns.
[0182] (3) In order to further improve the detection performance, an overall data-feature-decision fusion scheme was adopted in the entire anomaly detection process for time-frequency-space three-level fusion modeling, to replace the traditional shallow fusion strategy.
[0183] Extensive experiments conducted on eight publicly available industrial multi-sensor datasets demonstrate that the proposed time-frequency-space three-level fusion-driven anomaly detection method for industrial multi-sensor signals achieves excellent and stable detection accuracy. It consistently ranks among the top two in five out of the six evaluation metrics across all datasets. Furthermore, the method exhibits good efficiency, achieving reasonable inference speed and low resource consumption. Finally, a case study conducted on a physical test platform verifies the practicality of the time-frequency-space three-level fusion-driven detection method in real-world industrial scenarios.
[0184] The above embodiments are merely preferred examples of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.
Claims
1. A method for detecting anomalies in industrial multi-element sensor signals using time-frequency-space three-level fusion, characterized in that, Includes the following steps: Step S1: Acquire multi-dimensional time-series signals collected by multiple sensors during the industrial process; Step S2: For each timestamp, based on its historical time domain context, the frequency domain amplitude spectrum obtained by Fourier transform, and the spatial correlation pattern between multiple variables, a local pattern fourth-order tensor and a global pattern fourth-order tensor are constructed respectively to achieve data-level time-frequency-space fusion. Step S3: Perform orthogonal tensor decomposition on the local mode fourth-order tensor and the global mode fourth-order tensor to obtain multiple orthogonal two-dimensional matrices. Use a bidirectional large kernel convolutional network to extract time-domain features, frequency-domain features and spatial features from the orthogonal two-dimensional matrices. Then, concatenate and fuse the extracted features to achieve feature-level time-frequency-space fusion. Step S4: Based on the fused spatiotemporal features, calculate the contrast difference between the local and global patterns of each timestamp from the perspectives of intra-variable context and inter-variable coupling, respectively, and then weight and fuse the contrast differences from the two perspectives to generate the final anomaly score for each timestamp, thereby achieving decision-level spatiotemporal fusion. Step S5: Based on the comparison result between the final anomaly score and the preset threshold, determine whether the corresponding timestamp is an anomaly.
2. The industrial multi-sensor signal anomaly detection method based on time-frequency-space three-level fusion according to claim 1, characterized in that, Step S2 specifically includes the following: For each timestamp n, extract 2L consecutive historical timestamps forward from that timestamp to form a local context sequence; starting from timestamp n-2L, perform interval sampling at 2L steps, repeating 2L times to form a global context sequence; where 2L is the preset context length. Perform Fourier transform on the local context sequence and the global context sequence to extract the amplitude spectrum of length L for each timestamp; The timestamp dimension, sensor variable dimension, context sequence dimension, and amplitude spectrum dimension are stacked to construct the local mode fourth-order tensor and the global mode fourth-order tensor, respectively.
3. The industrial multi-sensor signal anomaly detection method based on time-frequency-space three-level fusion according to claim 1, characterized in that, The orthogonal tensor decomposition in step S3 uses Tucker decomposition, which decomposes the fourth-order tensor into three mutually orthogonal two-dimensional subspace matrices along the time dimension: time-context, time-space, and time-spectrum.
4. The industrial multi-sensor signal anomaly detection method based on time-frequency-space three-level fusion according to claim 1, characterized in that, The bidirectional large-kernel convolutional network in step S3 includes: The temporal dynamic learning branch employs a two-layer temporal convolutional network, utilizing large-size one-dimensional convolutional kernels to capture the temporal evolution features of local and global patterns along the temporal dimension. The pattern association learning branch uses full-scale and half-scale large-size one-dimensional convolutional kernels to capture the dependencies between feature values along the pattern dimension. The bidirectional feature fusion module is used to fuse the features output by the time-dynamic learning branch and the pattern association learning branch.
5. The industrial multi-sensor signal anomaly detection method based on time-frequency-space three-level fusion according to claim 4, characterized in that, In the temporal dynamic learning branch, the first layer of the two-layer temporal convolutional network uses a large-size one-dimensional convolutional kernel that matches the size of the context window to perform causal temporal convolution, and the second layer uses another large-size one-dimensional convolutional kernel to expand the temporal receptive field; in the pattern association learning branch, feature extraction is first performed using a full-scale large-kernel one-dimensional convolutional kernel with a length equal to the number of sensor variables M, and then further learning is performed using a half-scale large-kernel one-dimensional convolutional kernel with a length of M / 2.
6. The industrial multi-sensor signal anomaly detection method based on time-frequency-space three-level fusion according to claim 1, characterized in that, In step S4, the comparison difference calculation method of intra-variable context perspective is as follows: the intra-variable context is reconstructed by using the first multilayer perceptron to reconstruct the local time-frequency-space fusion features and the global time-frequency-space fusion features respectively, and the difference between the reconstruction results is calculated as the first comparison difference.
7. The industrial multi-sensor signal anomaly detection method based on time-frequency-space three-level fusion according to claim 1, characterized in that, In step S4, the comparison difference calculation method of the coupling perspective between variables is as follows: the coupling relationship between variables is reconstructed by using the second multilayer perceptron to reconstruct the local time-frequency-space fusion features and the global time-frequency-space fusion features respectively, and the difference between the reconstruction results is used as the second comparison difference.
8. The industrial multi-sensor signal anomaly detection method based on time-frequency-space three-level fusion according to claim 1, characterized in that, The weighted fusion in step S4 uses a balancing hyperparameter β to sum the contrast differences between the two perspectives. The value range of the hyperparameter β is [0,1].