Abnormality identification and fault prediction method of electromechanical equipment based on multimodal OOD detection

Abstract

This invention discloses a method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection, belonging to the field of electromechanical equipment detection. The method includes: S1, collecting time-series data from multimodal sensors of an electromechanical equipment cluster and representing it as a three-dimensional data structure; the multimodal sensor time-series data consists of ID data that has not experienced anomalies; S2, training an AE network using the ID data and freezing the parameters of the trained AE network; S3, constructing an OOD detector based on the AE network with frozen parameters; S4, using the OOD detector to identify anomalies in the input samples of the three-dimensional data structure, and locating abnormal equipment based on the reconstruction loss values ​​of each modal sensor in the equipment dimension, thereby achieving anomaly identification and fault prediction. This invention overcomes the limitations of traditional methods through multimodal data fusion and OOD detection technology, resulting in more accurate, robust, and interpretable fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of electromechanical equipment testing, specifically relating to a method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection. Background Technology

[0002] Currently, in the field of anomaly identification and fault prediction of electromechanical equipment, traditional methods face two key technical bottlenecks: first, fault location lacks interpretability; and second, the rate of missed detection for unknown faults remains high.

[0003] While existing technologies can detect faults based on single-mode data such as vibration, current, and temperature, the lack of an effective multi-mode feature fusion mechanism and an interpretable analysis framework results in the system only being able to output simple abnormal alarms, failing to clearly indicate the specific component where the fault occurred and its cause, which greatly complicates subsequent maintenance decisions.

[0004] More seriously, current mainstream methods generally adopt a closed-set learning paradigm, assuming that all fault types have already appeared during the training phase. This makes the system ineffective when facing entirely new fault modes that may appear in actual operation, resulting in a significantly higher false negative rate. This limitation is particularly prominent in industrial settings because the fault modes of electromechanical equipment have dynamic evolution characteristics and are affected by multiple factors such as environment and load. Methods that rely entirely on historical data for supervised learning are ill-suited to meet the fault detection needs in such open environments.

[0005] Although deep learning technology has shown advantages in automatic feature extraction in recent years, most models still operate at the "black box" level, unable to explain the physical mechanisms of faults or effectively identify abnormal patterns outside the training data distribution, which seriously restricts the promotion and application of intelligent diagnostic systems in actual industrial scenarios.

[0006] Currently, complex electromechanical systems involve a wide variety of equipment, complex operating states, and diverse fault types. Traditional fault diagnosis methods struggle to meet the requirements of high accuracy, robustness, and good interpretability. First, traditional supervised learning methods based on the closed-set assumption typically only identify a limited number of fault modes known during training. For out-of-distribution (OOD) samples—unknown faults not present in the training set—their identification ability is weak, resulting in a high false negative rate and serious safety hazards. Second, most existing systems can only provide simple alarms based on anomaly thresholds or predicted probabilities, failing to pinpoint the specific fault-causing equipment or component, let alone provide effective explanations of the fault's mechanism or characteristics. This limits their usability and reliability in real-world industrial scenarios. Summary of the Invention

[0007] To address the aforementioned shortcomings in existing technologies, the multimodal OOD detection method for electromechanical equipment anomaly identification and fault prediction provided by this invention solves the problems of lack of interpretability in fault location and high false negative rate for unknown faults when traditional methods are used to identify and predict anomalies in electromechanical equipment.

[0008] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection, comprising the following steps:

[0009] S1. Collect multimodal sensor time-series data of the electromechanical equipment cluster and represent it as a three-dimensional data structure;

[0010] The multimodal sensor timing data consists of ID data that has not experienced any anomalies.

[0011] S2. Train the AE network using the ID data and freeze the parameters of the trained AE network.

[0012] S3. Construct an OOD detector based on the AE network with frozen parameters;

[0013] S4. Use an OOD detector to identify anomalies in the input samples of the 3D data structure, and locate abnormal devices based on the reconstruction loss values ​​of each modal sensor in the device dimension, thereby realizing anomaly identification and fault prediction.

[0014] Further, in step S1, the three-dimensional data structure of the sensor time series data is represented as (N,D,S), where N is the number of samples, representing the number of observations collected within a set time or time period, D is the number of electromechanical devices, representing the total number of independent devices in the electromechanical device cluster, and S represents the number of sensor types configured in each device.

[0015] Furthermore, in step S2, the AE network includes an encoder and a decoder that maintain a symmetrical structure;

[0016] The specific method for training the AE network is as follows:

[0017] The input data is mapped to the latent space by the encoder, and the latent feature data is mapped back to the original space by the decoder to obtain the reconstructed data; wherein, the latent feature data in the latent space is fused with the multimodal feature information of the ID data;

[0018] The reconstruction error between the input ID data and the reconstructed data is used as the loss function to train the AE network.

[0019] Furthermore, the process of mapping the input data to the latent space through the encoder is specifically as follows:

[0020] The input data is linearly transformed into latent space data;

[0021] Based on the attention function learned from the latent space data, weights are assigned to each modality of the input data to obtain its corresponding assignable weights;

[0022] Normalized attention weights are obtained by applying a softmax operation to the assignable weights.

[0023] Attention weights are used to perform weighted summation of latent space data to obtain fused feature data that incorporates multimodal feature information;

[0024] The fused feature data is mapped to the latent space using a fully connected layer to obtain latent feature data.

[0025] Furthermore, the latent feature data Expressed as:

[0026]

[0027]

[0028]

[0029]

[0030]

[0031] In the formula, This indicates the fusion of feature data. This represents the weights that map the fused feature data to the latent space. Indicates the third bias term. This represents the ReLU activation function. This represents the normalized attention weight at time k in the d-th sensor of the n-th device. This represents the assignable weight of the d-th sensor in the n-th device. This represents the assignable weight at time k in the d-th sensor of the n-th device. This represents the assignable weight at time j in the d-th sensor of the n-th device. Represents latent space data. This represents the latent space data at time k in the d-th sensor of the n-th device. This represents the second-level projection vector. Represents a non-linear activation function. Indicates the second bias term. This represents the weights of the first-level linear transformation. Indicates the second bias term. This represents the input data, which is the sensor observation vector for each device n and sensor d. The superscript T indicates the transpose operation, and the subscripts j and k are time sequence numbers. .

[0032] Furthermore, in step S2, the loss function of the AE network is trained. for:

[0033]

[0034] In the formula, This represents the input data of the d-th sensor in the n-th device at time s. This represents the reconstructed data of the d-th sensor in the n-th device at time s. This indicates the sample size. This indicates the number of electromechanical devices, and S indicates the number of sensor types configured in each device.

[0035] Furthermore, in step S3, the OOD detector is used to evaluate the degree of deviation of the input sample from the ID data distribution, thereby realizing OOD data recognition;

[0036] The method for OOD data identification by the OOD detector is as follows:

[0037] Calculate the reconstruction loss value of the input sample after passing through the parameter-frozen AE network, compare the reconstruction loss value with the set discrimination threshold. If the reconstruction loss value exceeds the set discrimination threshold, the input sample is OOD data; otherwise, it is ID data.

[0038] The OOD data refers to data that has experienced an anomaly.

[0039] Furthermore, the method for determining the discrimination threshold is as follows:

[0040] For a training set containing only ID data, calculate its reconstruction loss value after passing through an AE network with frozen parameters;

[0041] The mean of the reconstruction loss values ​​for all ID data is calculated. and standard deviation Therefore, the discrimination threshold is determined. .

[0042] Furthermore, in step S4, the method for locating the abnormal device specifically includes:

[0043] For input samples identified as abnormal, the reconstruction loss values ​​of all sensor channels on each device are summed up one by one to calculate the total reconstruction loss value in that device dimension. The device with the largest total reconstruction loss value is identified as the abnormal device in the localization.

[0044] Furthermore, the total reconstruction loss value for:

[0045]

[0046] In the formula, This represents the input data of the d-th sensor in the n-th device at time s. This represents the reconstructed data of the d-th sensor in the n-th device at time s.

[0047] The beneficial effects of this invention are as follows:

[0048] This invention overcomes the limitations of traditional methods by combining multimodal data fusion and OOD detection technology, resulting in more accurate, robust, and interpretable fault diagnosis. Specifically, it is reflected in:

[0049] (1) Unlike traditional methods that rely on single-modal data, this invention integrates multimodal data such as vibration, temperature, and current, and represents them in the form of a three-dimensional data structure. Through this multi-dimensional time-series structure, not only can the operating status of a single device be tracked, but also the correlation patterns and fault propagation paths between devices in the cluster can be analyzed from a global perspective. This enhances the stability and accuracy of the system under missing or noise interference conditions, and provides a solid data foundation for subsequent intelligent operation and maintenance, predictive maintenance and risk warning.

[0050] (2) The present invention uses the unsupervised learning method of AE network to effectively identify unknown fault modes, and uses a data fusion method based on modal attention mechanism to fuse multiple sensor data. By setting the discrimination threshold based on the reconstruction loss value, it can accurately distinguish normal data from abnormal data, thereby improving the system's adaptability to new faults.

[0051] (3) In terms of fault location, the present invention can accurately locate faulty equipment by analyzing equipment-level reconstruction error. It can effectively fuse abnormal signals at the modal level, thereby enhancing the robustness and reliability of fault location and avoiding the global alarm problem of traditional methods.

[0052] (4) The present invention automatically sets a reasonable discrimination threshold through statistical methods. This method does not rely on any out-of-distribution data or label information, and is therefore particularly suitable for unsupervised or semi-supervised scenarios. On the other hand, the threshold mechanism constructed by statistics also enhances the stability and robustness of the detection process. Because it utilizes the quantitative relationship between the central tendency of the distribution and the degree of outlier, it can effectively distinguish normal samples from potential abnormal samples, maintain consistent detection performance in different datasets or application scenarios, and reduce false alarms.

[0053] Overall, this invention demonstrates higher accuracy and robustness in fault detection and location, providing strong support for equipment maintenance and management. Attached Figure Description

[0054] Figure 1 The flowchart of the electromechanical equipment anomaly identification and fault prediction method based on multimodal OOD detection provided by the present invention is shown. Detailed Implementation

[0055] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0056] This invention proposes a method for anomaly identification and fault prediction that integrates multimodal information and an OOD (Out-of-Depth) detection mechanism, to achieve comprehensive perception of the operating status of electromechanical equipment, accurate fault identification, and interpretable localization of the fault's location and cause. Specifically, the electromechanical equipment anomaly identification and fault prediction method based on multimodal OOD detection in this invention, such as... Figure 1 As shown, it includes the following steps:

[0057] S1. Collect multimodal sensor time-series data of the electromechanical equipment cluster and represent it as a three-dimensional data structure;

[0058] The multimodal sensor timing data consists of ID data that has not experienced any anomalies.

[0059] S2. Train the AE network using the ID data and freeze the parameters of the trained AE network.

[0060] S3. Construct an OOD detector based on the AE network with frozen parameters;

[0061] S4. Use an OOD detector to identify anomalies in the input samples of the 3D data structure, and locate abnormal devices based on the reconstruction loss values ​​of each modal sensor in the device dimension, thereby realizing anomaly identification and fault prediction.

[0062] In step S1 of this embodiment of the invention, the three-dimensional data structure of the sensor time series data is represented as (N,D,S), where N is the number of samples, representing the number of observations collected within a set time or time period, D is the number of electromechanical devices, representing the total number of independent devices in the electromechanical device cluster, and S represents the number of sensor types configured in each device.

[0063] In this embodiment, a multi-dimensional time-series organization method for multi-electromechanical equipment clusters is proposed. This method effectively expresses and manages the complex data structure of multiple devices, multiple sensors, and multiple time points in a large-scale electromechanical system, which is represented as a three-dimensional data structure (N,D,S). Here, N represents the number of samples, specifically the number of observations collected at a certain moment or within a certain time period; D represents the number of electromechanical equipment, that is, the total number of independent devices in the cluster; and S represents the number of sensor types configured in each device, such as common temperature, vibration, current, and other types of sensors.

[0064] Specifically, this three-dimensional data structure possesses excellent scalability and organization, clearly capturing data information at different levels (time, device, sensor) of the system, enabling structured data representation and efficient access. Compared to traditional two-dimensional tables or flat storage methods, the three-dimensional structure not only reduces data redundancy and improves storage and processing efficiency, but also facilitates subsequent applications in tasks such as machine learning, state monitoring, and anomaly detection. For example, when using deep learning models for modeling and analysis, this data structure is naturally adapted to multi-dimensional tensor inputs, fully preserving the correlation of the original data in the temporal, spatial, and functional dimensions, thereby enhancing the model's understanding and predictive ability of system behavior.

[0065] Furthermore, in complex industrial environments where multiple devices operate collaboratively, there may be dynamic characteristics of coupling or mutual influence between different devices. This three-dimensional data structure not only allows for tracking the operational status of individual devices but also enables analysis of the correlation patterns and fault propagation paths between devices in a cluster from a global perspective, providing a solid data foundation for subsequent intelligent operation and maintenance, predictive maintenance, and risk warning.

[0066] In step S2 of this embodiment, in order to effectively integrate multimodal data collected by various sensors (such as temperature, vibration, current, etc.) in electromechanical equipment, this invention proposes an AE network with a fusion modal attention mechanism. This network models the ID data of the normal operating state (In-Distribution, ID) and achieves accurate detection of OOD data (out-of-distribution data) through reconstruction error. When the reconstruction error of a test sample significantly deviates from the training distribution, it can be determined as a possible unknown fault.

[0067] The AE network in this embodiment of the invention includes an encoder and a decoder that maintain a symmetrical structure;

[0068] The specific method for training the AE network is as follows:

[0069] The input data is mapped to the latent space by the encoder, and the latent feature data is mapped back to the original space by the decoder to obtain the reconstructed data; wherein, the latent feature data in the latent space is fused with the multimodal feature information of the ID data;

[0070] The reconstruction error between the input ID data and the reconstructed data is used as the loss function to train the AE network.

[0071] Specifically, based on the above AE network structure, the process of mapping input data to the latent space through the encoder is as follows:

[0072] The input data is linearly transformed into latent space data;

[0073] Based on the attention function learned from the latent space data, weights are assigned to each modality of the input data to obtain its corresponding assignable weights;

[0074] Normalized attention weights are obtained by applying a softmax operation to the assignable weights.

[0075] Attention weights are used to perform weighted summation of latent space data to obtain fused feature data that incorporates multimodal feature information;

[0076] The fused feature data is mapped to the latent space using a fully connected layer to obtain latent feature data.

[0077] Specifically, under sensor d for each device n, the sensor observation vector The original data is linearly transformed into more discriminative latent space data.

[0078]

[0079] Based on tacit space data Computational attention function for learning Assign weights to each modality; weights can be assigned. for:

[0080]

[0081] Next, a softmax operation is applied to the assignable weights to obtain normalized attention weights. for:

[0082]

[0083] Using attention weights We obtain fused feature data by weighted summation of the latent space data. for:

[0084]

[0085] Through the aforementioned modal attention mechanism, the AE network can adaptively capture the importance of different sensors under specific time and device conditions, thereby generating a more discriminative fusion representation. The fused feature data is mapped to the latent space using a fully connected layer to obtain the latent feature data. for:

[0086]

[0087] In the formula, This indicates the fusion of feature data. This represents the weights that map the fused feature data to the latent space. Indicates the third bias term. This represents the ReLU activation function. This represents the normalized attention weight at time k in the d-th sensor of the n-th device. This represents the assignable weight of the d-th sensor in the n-th device. This represents the assignable weight at time k in the d-th sensor of the n-th device. This represents the assignable weight at time j in the d-th sensor of the n-th device. Represents latent space data. This represents the latent space data at time k in the d-th sensor of the n-th device. This represents the second-level projection vector. Represents a non-linear activation function. Indicates the second bias term. This represents the weights of the first-level linear transformation. Indicates the second bias term. This represents the input data, which is the sensor observation vector for each device n and sensor d. The superscript T indicates the transpose operation, and the subscripts j and k are time sequence numbers. , where m represents the potential spatial dimension.

[0088] Based on the latent feature data obtained by the encoder, the decoder maps it back to the original space. .in, This indicates the decoder of the AE network.

[0089] In step S2 of this embodiment of the invention, the loss function of the AE network is trained. for:

[0090]

[0091] In the formula, This represents the input data of the d-th sensor in the n-th device at time s. This represents the reconstructed data of the d-th sensor in the n-th device at time s. This indicates the sample size. This indicates the number of electromechanical devices, and S indicates the number of sensor types configured in each device.

[0092] In step S3 of this embodiment of the invention, the OOD detector is used to evaluate the degree of deviation of the input sample from the ID data distribution, thereby realizing OOD data recognition; specifically, the method of OOD data recognition by the OOD detector is as follows:

[0093] Calculate the reconstruction loss value of the input sample after passing through the parameter-frozen AE network, compare the reconstruction loss value with the set discrimination threshold. If the reconstruction loss value exceeds the set discrimination threshold, the input sample is OOD data; otherwise, it is ID data.

[0094] The OOD data refers to data that has experienced an anomaly.

[0095] In this embodiment, a method based on statistical features is used to determine the discrimination threshold. The specific method for determining the discrimination threshold is as follows:

[0096] For a training set containing only ID data, calculate its reconstruction loss value after passing through an AE network with frozen parameters;

[0097] The mean of the reconstruction loss values ​​for all ID data is calculated. and standard deviation Therefore, the discrimination threshold is determined. .

[0098] In this embodiment, the discrimination threshold determined by the above-mentioned "three standard deviation principle" method can accurately distinguish between data within and outside the distribution, thereby achieving high detection capability while maintaining a low false alarm rate.

[0099] The "three standard deviations rule," as a method for identifying outliers, is based on the idea that, under the assumption of a normal distribution, the vast majority (approximately 99.7%) of the data will fall within the range of the mean plus or minus three standard deviations. Therefore, when training an AE network using only ID samples, its reconstruction error should statistically be concentrated within this range. Thus, if the reconstruction error of a test sample is significantly higher than this range, it is highly likely that its feature distribution differs significantly from the training data, and can therefore be reasonably identified as an OOD sample.

[0100] Furthermore, this method exhibits good versatility and interpretability in practical applications. On one hand, since it does not rely on any out-of-distribution data or label information, it is particularly suitable for unsupervised or semi-supervised scenarios. On the other hand, the threshold mechanism constructed using statistics enhances the stability and robustness of the detection process, enabling consistent detection performance across different datasets or application scenarios. It is worth noting that although the standard deviation principle theoretically assumes that the reconstruction error follows a normal distribution, in practical applications, even if the error distribution does not perfectly conform to normality, this method still achieves good approximation results because it utilizes the quantitative relationship between the central tendency of the distribution and the degree of outlier, thereby effectively distinguishing normal samples from potentially anomalous samples.

[0101] In summary, this embodiment adopts a method based on " The threshold setting method of "" not only has a clear theoretical basis and is easy to implement, but also shows strong practical value under the AE network architecture, and is an effective strategy in unsupervised anomaly detection.

[0102] In step S4 of this embodiment of the invention, the method for locating abnormal devices is specifically as follows:

[0103] For input samples identified as abnormal, the reconstruction loss values ​​of all sensor channels on each device are summed up one by one to calculate the total reconstruction loss value in that device dimension. The device with the largest total reconstruction loss value is identified as the abnormal device in the localization.

[0104] Among them, the total reconstruction loss value for:

[0105]

[0106] In the formula, in the formula, This represents the input data of the d-th sensor in the n-th device at time s. This represents the reconstructed data of the d-th sensor in the n-th device at time s. The formula means that, across all device dimensions, the device with the largest sum of reconstruction errors (i.e., the squared difference between the original and reconstructed values) is selected as the most likely target device in the current sample to be prone to anomalies.

[0107] Specifically, in this embodiment, based on the reconstruction error detection result, when the input sample at a certain moment... When an anomaly is identified, it indicates that one or more electromechanical devices in the current system may be malfunctioning. However, the detection results at this point only provide global information that an anomaly exists, and cannot pinpoint the specific device to which the fault originates. To further improve the diagnostic accuracy and localization capabilities of the system, a detailed analysis of the reconstruction errors across various device dimensions within the sample is needed to determine the most likely device to be malfunctioning.

[0108] For each input sample that is judged as abnormal Each device All sensor channels on The reconstruction errors are accumulated item by item, and the total reconstruction loss in the device dimension is calculated. The device with the largest total reconstruction loss value is regarded as the abnormal device in the localization.

[0109] The core idea of ​​the above-mentioned anomaly localization method is that when a certain electromechanical equipment has a potential fault, its sensor features often exhibit characteristics that are significantly different from the normal distribution in one or more modes, which leads to a significant increase in the reconstruction error in the corresponding dimension. By aggregating the reconstruction loss values ​​of all modes in the device dimension, the abnormal signals at the mode level can be effectively fused, thereby enhancing the robustness and reliability of anomaly localization.

[0110] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0111] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection, characterized in that, Includes the following steps: S1. Collect multimodal sensor time-series data of the electromechanical equipment cluster and represent it as a three-dimensional data structure; The multimodal sensor timing data consists of ID data that has not experienced any anomalies. S2. Train the AE network using the ID data and freeze the parameters of the trained AE network. S3. Construct an OOD detector based on the AE network with frozen parameters; S4. Use an OOD detector to identify anomalies in the input samples of the 3D data structure, and locate abnormal devices based on the reconstruction loss values ​​of each modal sensor in the device dimension, so as to realize anomaly identification and fault prediction. In step S3, the OOD detector is used to evaluate the degree of deviation of the input sample from the ID data distribution, thereby realizing OOD data recognition. The method for OOD data identification by the OOD detector is as follows: Calculate the reconstruction loss value of the input sample after passing through the parameter-frozen AE network, compare the reconstruction loss value with the set discrimination threshold. If the reconstruction loss value exceeds the set discrimination threshold, the input sample is OOD data; otherwise, it is ID data. The OOD data refers to data that has experienced an anomaly. The specific method for determining the discrimination threshold is as follows: For a training set containing only ID data, calculate its reconstruction loss value after passing through an AE network with frozen parameters; The mean of the reconstruction loss values ​​for all ID data is calculated. and standard deviation Therefore, the discrimination threshold is determined. .

2. The method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection according to claim 1, characterized in that, In step S1, the three-dimensional data structure of the sensor time series data is represented as (N,D,S), where N is the number of samples, representing the number of observations collected within a set time or time period, D is the number of electromechanical devices, representing the total number of independent devices in the electromechanical device cluster, and S represents the number of sensor types configured in each device.

3. The method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection according to claim 1, characterized in that, In step S2, the AE network includes an encoder and a decoder that maintain a symmetrical structure. The specific method for training the AE network is as follows: The input data is mapped to the latent space by the encoder, and the latent feature data is mapped back to the original space by the decoder to obtain the reconstructed data; wherein, the latent feature data in the latent space is fused with the multimodal feature information of the ID data; The reconstruction error between the input ID data and the reconstructed data is used as the loss function to train the AE network.

4. The method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection according to claim 3, characterized in that, The process of mapping input data to the latent space using an encoder is as follows: The input data is linearly transformed into latent space data; Based on the attention function learned from the latent space data, weights are assigned to each modality of the input data to obtain its corresponding assignable weights; Normalized attention weights are obtained by applying a softmax operation to the assignable weights. Attention weights are used to perform weighted summation of latent space data to obtain fused feature data that incorporates multimodal feature information; The fused feature data is mapped to the latent space using a fully connected layer to obtain latent feature data.

5. The method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection according to claim 4, characterized in that, The latent feature data Expressed as: In the formula, This indicates the fusion of feature data. This represents the weights that map the fused feature data to the latent space. Indicates the third bias term. This represents the ReLU activation function. This represents the normalized attention weight at time k in the d-th sensor of the n-th device. This represents the assignable weight of the d-th sensor in the n-th device. This represents the assignable weight at time k in the d-th sensor of the n-th device. This represents the assignable weight at time j in the d-th sensor of the n-th device. Represents latent space data. This represents the latent space data at time k in the d-th sensor of the n-th device. This represents the second-level projection vector. Represents a non-linear activation function. Indicates the second bias term. This represents the weights of the first-level linear transformation. Indicates the second bias term. This represents the input data, which is the sensor observation vector for each device n and sensor d. The superscript T indicates the transpose operation, and the subscripts j and k are time sequence numbers. .

6. The method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection according to claim 3, characterized in that, In step S2, the loss function of the AE network is trained. for: In the formula, This represents the input data of the d-th sensor in the n-th device at time s. This represents the reconstructed data of the d-th sensor in the n-th device at time s. This indicates the sample size. This indicates the number of electromechanical devices, and S indicates the number of sensor types configured in each device.

7. The method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection according to claim 1, characterized in that, In step S4, the method for locating abnormal devices is as follows: For input samples identified as abnormal, the reconstruction loss values ​​of all sensor channels on each device are summed up one by one to calculate the total reconstruction loss value in that device dimension. The device with the largest total reconstruction loss value is identified as the abnormal device in the localization.

8. The method for anomaly identification and fault prediction of electromechanical equipment based on multimodal OOD detection according to claim 7, characterized in that, The total reconstruction loss value for: In the formula, This represents the input data of the d-th sensor in the n-th device at time s. This represents the reconstructed data of the d-th sensor in the n-th device at time s.

Images