Multi-mode adaptive fault diagnosis methods, devices, equipment and storage media

Abstract

This application provides a multi-mode adaptive fault diagnosis method, apparatus, device, and storage medium, relating to the field of fault prediction technology. For each operating mode, a reversible mapping model is determined based on the corresponding healthy samples in the sample library. A unique fault sample for one operating mode is obtained. The unique fault sample is then remapped across operating conditions using at least two corresponding reversible mapping models to obtain a target mode fault sample. The sample library is updated based on the target mode fault sample. The updated sample library is used to train the initial fault diagnosis model, resulting in a trained fault diagnosis model. The distribution alignment of different modes in the latent space can be completed using only healthy samples. Cross-operating condition feature remapping transfers unique fault samples from a single mode to other modes, demonstrating good cross-operating condition generalization ability, adapting to mode switching scenarios under different operating conditions, and improving the fault diagnosis accuracy of multi-mode complex process systems.

Description

Technical Field

[0001] This application relates to the field of fault prediction technology, and in particular to multi-mode adaptive fault diagnosis methods, apparatus, devices and storage media. Background Technology

[0002] Complex process systems, such as nuclear power plant reactor cooling systems, are characterized by strong coupling of multiple variables, high nonlinearity, and multi-mode switching operation. In actual operation, the system switches operating modes according to changes in power load, fuel ratio, ambient temperature, and other conditions, and the distribution of process parameter data differs significantly between different modes.

[0003] Domain-adaptive fault diagnosis methods in related technologies typically employ techniques such as adversarial training, maximum mean difference alignment, and data augmentation to learn shared feature representations across operating conditions. By reducing the distributional differences of normal samples under different operating conditions, the classifier trained in the source domain is generalized to the target domain. However, these methods assume that fault categories completely overlap across all operating modes, which is inconsistent with real-world industrial scenarios. In complex process systems, some fault types only appear under specific operating modes, with no corresponding fault samples in other modes. In this scenario, domain alignment methods may incorrectly preserve fault classification boundaries in modes without corresponding faults, misclassifying normal operating data as faults, generating numerous false alarms, and directly leading to a decrease in fault diagnosis accuracy. Summary of the Invention

[0004] The main objective of this application is to propose a multi-mode adaptive fault diagnosis method, apparatus, device, and storage medium to improve the accuracy of adaptive fault diagnosis for complex process systems.

[0005] To achieve the above objectives, a first aspect of this application proposes a multi-mode adaptive fault diagnosis method, comprising: For each operating mode, a reversible mapping model is determined based on the corresponding healthy samples in the sample library, and all the reversible mapping models share the same latent space; Obtain a unique fault sample of one of the operating modes, obtain the initial reversible mapping model corresponding to the initial mode to which the unique fault sample belongs, and calculate the latent space vector corresponding to the unique fault sample in the initial reversible mapping model; obtain the target reversible mapping model corresponding to the target mode, and use the inverse mapping function of the target reversible mapping model to restore the latent space vector to the variable space of the target mode to obtain the corresponding target mode fault sample; update the sample library based on the target mode fault sample. The initial fault diagnosis model is trained using the updated sample library until a trained fault diagnosis model is obtained. The fault diagnosis model is used to perform online fault prediction on the online acquired data to be detected, and obtain fault diagnosis results.

[0006] In some embodiments, determining the reversible mapping model based on the corresponding healthy samples in the sample library includes: Determine the set of healthy samples corresponding to each of the operating modes from the sample library; Based on the reversible neural network, a reversible mapping model is constructed for each of the operating modes. The set of healthy samples is input into the corresponding reversible mapping model to obtain the mapping distribution of the set of healthy samples in the same latent space. With the objective of minimizing the maximum mean difference between any two of the said mapping distributions, the parameters of all the said invertible mapping models are adjusted to obtain the trained invertible mapping model.

[0007] In some embodiments, updating the sample library based on the target mode fault samples includes: Obtain any one of the healthy samples in the target mode, and use interpolation coefficients to interpolate the hidden representations corresponding to the healthy sample and the faulty sample of the target mode to obtain a synthetic hidden representation; The synthetic implicit representation is restored to the variable space of the target pattern using the inverse mapping model of the target invertible mapping model to obtain the synthetic sample; The process violation degree of the synthesized sample is obtained. If the process violation degree meets the preset threshold condition, the target mode fault sample is added to the sample library. Otherwise, the interpolation coefficient is resampled and the test is repeated within the retry limit until the preset threshold condition is met or the target mode fault sample is abandoned.

[0008] In some embodiments, the fault diagnosis model includes at least a dynamic pattern decoupling normalization layer, a multi-scale spatiotemporal feature extraction layer, a dual-stream collaborative focusing layer, and a classification decision layer. The multi-scale spatiotemporal feature extraction layer includes multiple parallel deep convolutional layers with different kernel sizes. The fault diagnosis model is used to perform the following steps when performing online fault prediction on online acquired data to be detected: The input features corresponding to the data to be detected are obtained, and the input features are input into the dynamic mode decoupling normalization layer for dynamic normalization to obtain a normalized output vector. The normalized output vector is input in parallel into each of the deep convolutional layers to extract local temporal features at different scales, and then concatenated along the channel dimension to obtain a multi-scale temporal feature vector. The multi-scale temporal feature vector is input into the dual-stream collaborative focusing layer for spatiotemporal joint attention calculation to obtain enhanced features; The classification decision layer is used to predict the probability of fault categories based on the enhanced features, thereby obtaining fault diagnosis results.

[0009] In some embodiments, the step of inputting the input features into the dynamic pattern decoupling normalization layer for dynamic normalization to obtain a normalized output vector includes: After performing global average pooling on the input features, the input pattern prediction network makes predictions to obtain the posterior probability of the output vector for each of the operating modes. Obtain the learnable mean and standard deviation parameters corresponding to each of the operating modes, and perform a weighted summation of the mean and standard deviation parameters based on the posterior probability to obtain the weighted mean and weighted standard deviation of the current sample. The input features are centered and scaled using the weighted mean and the weighted standard deviation to obtain the normalized output vector.

[0010] In some embodiments, the step of inputting the multi-scale temporal feature vector into the dual-stream co-focusing layer for spatiotemporal joint attention calculation to obtain enhanced features includes: The multi-scale temporal feature vectors are adaptively pooled along the channel direction to obtain spatial attention vectors, and adaptively pooled along the time direction to obtain temporal attention vectors. The spatiotemporal joint attention vector is obtained based on the temporal attention vector and the spatial attention vector. The spatiotemporal joint attention vector and the multi-scale temporal feature vector are multiplied element-wise and summed according to the time dimension to obtain the enhanced feature.

[0011] In some embodiments, the method further includes: Obtain the confidence level of the fault diagnosis result output by the fault diagnosis model after performing online fault prediction on the data to be detected; If the confidence level is greater than or equal to the preset confidence level threshold, the data to be detected and the fault diagnosis result are combined into a new sample with a pseudo-label and stored in the cache pool. The fault diagnosis model is then fine-tuned in a small batch using the new sample in the cache pool.

[0012] In some embodiments, the step of using the new samples in the cache pool to perform mini-batch fine-tuning of the fault diagnosis model includes: When the number of new samples accumulated in the cache pool reaches a preset threshold, the total loss function is the sum of cross-entropy loss and regularization loss. The fault diagnosis model is then fine-tuned using the new samples in the cache pool in a small batch. After the fine-tuning is completed, the cache pool is cleared.

[0013] To achieve the above objectives, a second aspect of this application provides a multi-mode adaptive fault diagnosis device, comprising: Reversible mapping module: used to determine a reversible mapping model for each operating mode based on the corresponding healthy samples in the sample library, and all the reversible mapping models share the same latent space; Feature remapping module: used to obtain a unique fault sample of one of the operating modes, obtain the initial reversible mapping model corresponding to the initial mode to which the unique fault sample belongs, calculate the latent space vector corresponding to the unique fault sample in the initial reversible mapping model; obtain the target reversible mapping model corresponding to the target mode, use the inverse mapping function of the target reversible mapping model to restore the latent space vector to the variable space of the target mode, obtain the corresponding target mode fault sample, and update the sample library based on the target mode fault sample; Fault diagnosis training module: used to train the initial fault diagnosis model using the updated sample library until a trained fault diagnosis model is obtained. The fault diagnosis model is used to perform online fault prediction on the online acquired data to be detected, and obtain fault diagnosis results.

[0014] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0015] To achieve the above objectives, a fourth aspect of the present application provides a storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0016] The multi-mode adaptive fault diagnosis method, apparatus, device, and storage medium proposed in this application, for each operating mode, determines a reversible mapping model based on the corresponding healthy samples in the sample library, with all reversible mapping models sharing the same latent space; obtains a unique fault sample for one operating mode, obtains the initial reversible mapping model corresponding to the initial mode to which the unique fault sample belongs, and calculates the latent space vector corresponding to the unique fault sample in the initial reversible mapping model; obtains the target reversible mapping model corresponding to the target mode, and uses the inverse mapping function of the target reversible mapping model to restore the latent space vector to the variable space of the target mode to obtain the corresponding target mode fault sample; updates the sample library based on the target mode fault sample; and trains the initial fault diagnosis model using the updated sample library until a trained fault diagnosis model is obtained. The fault diagnosis model is used to perform online fault prediction on the online acquired data to be detected, and obtains the fault diagnosis result. This application embodiment constructs reversible mapping models sharing the same latent space using healthy samples corresponding to each operating mode, and the distribution alignment of different modes in the latent space can be completed using only healthy samples. Based on this, unique fault samples from a single mode are transferred to other target modes through cross-condition feature remapping, generating corresponding fault samples for the target modes and expanding the sample library to fill the gaps in the corresponding fault samples for the target modes. The fault diagnosis model trained based on the expanded sample library has good cross-condition generalization ability, can adapt to mode switching scenarios under different loads and operating conditions, and improves the fault diagnosis accuracy of multi-mode complex process systems. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-mode adaptive fault diagnosis method provided in the embodiments of this application.

[0018] Figure 2 This is a flowchart of determining a reversible mapping model based on the corresponding healthy samples in the sample library, provided in an embodiment of this application.

[0019] Figure 3 This is a flowchart illustrating the process of updating the sample library based on target mode fault samples, as provided in an embodiment of this application.

[0020] Figure 4 This is a schematic diagram of the sample expansion process provided in this embodiment.

[0021] Figure 5 This is a flowchart illustrating how the fault diagnosis model provided in this application performs online fault prediction on the online-acquired data to be tested.

[0022] Figure 6 This is a flowchart provided in this application embodiment of the process of decoupling the input feature input dynamic pattern normalization layer and performing dynamic normalization to obtain a normalized output vector.

[0023] Figure 7 This is a schematic diagram of the processing flow of the dynamic mode decoupling normalization layer provided in the embodiments of this application.

[0024] Figure 8 This is a flowchart provided in an embodiment of the present application, in which multi-scale temporal feature vectors are input into a dual-stream collaborative focusing layer for spatiotemporal joint attention calculation to obtain enhanced features.

[0025] Figure 9 This is a schematic diagram of the calculation and processing flow of the enhanced features provided in the embodiments of this application.

[0026] Figure 10 This is a flowchart of the adaptive update of the fault diagnosis model provided in the embodiments of this application.

[0027] Figure 11 This is an overall flowchart of the multi-mode adaptive fault diagnosis provided in the embodiments of this application.

[0028] Figure 12 This is a structural block diagram of a multi-mode adaptive fault diagnosis device provided in another embodiment of this application.

[0029] Figure 13 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0031] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0033] Complex process systems, such as nuclear power plant reactor cooling systems, are characterized by multivariability, strong coupling, nonlinearity, and multi-mode operation. In actual production, the system switches operating modes under different loads, different feedstock or fuel ratios, and different ambient temperatures, resulting in significant differences in the distribution of process data across modes. In particular, certain types of faults (such as coking caused by specific feedstocks, abnormal vibration under specific loads, and sensor drift at specific power levels in nuclear power plants) may only occur in some modes, while corresponding fault samples cannot be obtained in other modes.

[0034] Fault diagnosis in related technologies includes single-mode deep learning diagnostic methods, domain adaptation and domain generalization methods, physical model-based diagnostic methods, and ensemble learning and transfer learning methods. Single-mode deep learning diagnostic methods employ convolutional neural networks, long short-term memory networks, and gated recurrent units to train an end-to-end classifier using a large amount of labeled data under a single operating mode. Domain adaptation and domain generalization methods learn cross-mode shared feature representations through adversarial training (such as domain adversarial neural networks), maximum mean difference alignment, and data augmentation techniques. Representative works include deep domain adaptive networks and invariant risk minimization, aiming to reduce the distribution differences of normal samples under different operating conditions, thereby generalizing the classifier trained in the source domain to the target domain. Physical model-based diagnostic methods utilize Kalman filtering, state observers, and mechanistic model residual analysis to diagnose faults by comparing the deviation between actual output and model prediction. Typical applications include reactor condition monitoring in nuclear power plants and compressor gas path diagnosis. Ensemble learning and transfer learning methods employ strategies such as multi-source domain transfer learning, meta-learning, and model ensemble to train multiple base learners on multiple source scenarios, and then adapt them to the target scenario through weighting or selection mechanisms. Some works introduce temporal attention mechanisms to capture slowly drifting features.

[0035] However, single-mode deep learning diagnostic methods are completely unable to adapt to the data distribution drift caused by mode switching. When the process system switches between different loads and different raw material ratios, the model accuracy drops significantly. It lacks both the ability to generalize to unknown operating conditions and the ability to handle unique fault types that only appear in certain operating modes. Domain adaptation and domain generalization methods generally assume that fault categories completely overlap in each operating mode. However, in real-world complex process systems, some faults only appear in specific operating modes, and there are no corresponding fault samples in other modes. In such scenarios, domain alignment methods incorrectly retain fault classification boundaries in modes without corresponding faults, misclassifying normal operating data as faults and triggering numerous false alarms. Furthermore, these methods do not consider gradual process changes such as catalyst deactivation and heat exchanger fouling, resulting in insufficient detection capabilities for early, subtle faults. Diagnostic methods based on physical models are limited by the strong nonlinearity, strong coupling, and incomplete mechanisms of complex process systems, making it difficult to establish accurate analytical models. Model simplification assumptions introduce non-negligible modeling errors, making it difficult to set residual thresholds, resulting in persistently high false alarm and false negative rates. Meanwhile, physical models struggle to incorporate data-driven adaptive update capabilities, making them ill-equipped to handle long-term process drift. Ensemble learning and transfer learning methods suffer from high computational overhead, and multi-model integration leads to significant inference latency, failing to meet real-time diagnostic requirements. Furthermore, these methods lack specialized modeling for gradual process changes, resulting in low sensitivity to early, minor faults. During long-term system operation, the model cannot utilize unlabeled real-time data for progressive updates, leading to a continuous decline in diagnostic performance as the process drifts.

[0036] The design logic of these methods in related technologies cannot simultaneously address the three major problems of incomplete overlap of fault categories, modeling of gradual process changes, and long-term drift adaptation. Specifically, multi-mode domain alignment methods assume all domains share the same fault space, which is inconsistent with real-world industrial scenarios; single-mode models lack cross-condition transfer capabilities; physical models struggle to handle system nonlinearity and uncertainty; and transfer learning ignores proactive adaptation to gradual process changes. From a fundamental design perspective, these solutions assume that fault categories are completely consistent across all operating conditions, and that data distribution drift can be eliminated through simple domain alignment. However, in actual industrial scenarios, fault samples naturally exist in different operating modes, compounded by long-term gradual processes such as slow catalyst deactivation and heat exchanger fouling, directly causing traditional deep transfer learning methods to fail.

[0037] Meanwhile, as critical infrastructure such as nuclear power plants becomes increasingly digitalized and networked, industrial control systems face extreme risks of coupled network attacks and physical faults, making diagnostic methods based on single-mode or full-fault-category assumptions inapplicable. Therefore, a novel fault diagnosis architecture is needed that breaks the assumption of full fault category overlap, integrates prior process knowledge, and combines gradual process modeling with long-term drift adaptation capabilities.

[0038] Based on this, embodiments of this application provide a multi-mode adaptive fault diagnosis method, apparatus, device, and storage medium. By constructing reversible mapping models sharing the same latent space using health samples corresponding to each operating mode, the distribution alignment of different modes in the latent space can be achieved using only health samples. Furthermore, through cross-condition feature remapping, unique fault samples from a single mode are transferred to other target modes, generating fault samples corresponding to the target modes and expanding the sample library to fill the gaps in corresponding fault samples for the target modes. The fault diagnosis model trained based on the expanded sample library possesses good cross-condition generalization ability, adapting to mode switching scenarios under different loads and operating conditions, thus improving the fault diagnosis accuracy of multi-mode complex process systems.

[0039] This application provides a multi-mode adaptive fault diagnosis method, apparatus, device, and storage medium, which are specifically described through the following embodiments. First, the multi-mode adaptive fault diagnosis method in this application is described.

[0040] This application's embodiments can acquire and process relevant data based on artificial intelligence (AI) technology. AI is the theory, methods, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new type of intelligent machine that can react in a way similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0041] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0042] The multi-mode adaptive fault diagnosis method provided in this application relates to the field of fault prediction technology. This method can be applied to a terminal, a server, or a computer program running on either the terminal or the server. For example, the computer program can be a native program or software module in an operating system; it can be a native application (APP), i.e., a program that needs to be installed in the operating system to run, such as a client supporting multi-mode adaptive fault diagnosis, i.e., a program that only needs to be downloaded to a browser environment to run; or it can be a small program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of application, module, or plugin. The terminal communicates with the server via a network. This multi-mode adaptive fault diagnosis method can be executed by the terminal or the server, or by the terminal and the server working together.

[0043] In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, or smartwatch, etc. The server can be a standalone server, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms; it can also be a service node in a blockchain system, where the service nodes form a peer-to-peer (P2P) network. The P2P protocol is an application layer protocol running on top of the Transmission Control Protocol (TCP). The terminal and server can connect via Bluetooth, Universal Serial Bus (USB), or a network, etc., and this embodiment does not impose any limitations.

[0044] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0045] The multi-mode adaptive fault diagnosis method in the embodiments of this application is described below.

[0046] Figure 1 This is an optional flowchart of the multi-mode adaptive fault diagnosis method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps 110 to 130. It is also understood that this embodiment... Figure 1 The order of steps 110 to 130 is not specifically limited. The order of steps can be adjusted or some steps can be reduced or added according to actual needs.

[0047] Step 110: For each operating mode, determine the reversible mapping model based on the corresponding healthy samples in the sample library.

[0048] In one embodiment, a bidirectional mapping capability is established from healthy samples of different operating modes to a unified feature space, facilitating cross-mode migration of unique fault samples from certain modes. All reversible mapping models share the same latent space, thereby ensuring that healthy samples from different operating modes can achieve distribution alignment under a unified feature dimension. (Refer to...) Figure 2 , Figure 2 This is a flowchart provided in this application embodiment for determining a reversible mapping model based on corresponding healthy samples in a sample library, specifically including the following steps: Step 210: Determine the set of healthy samples corresponding to each operating mode from the sample library.

[0049] In one embodiment, the sample library stores historical data of a complex process system across all operating conditions. This historical data includes healthy samples, which are multivariate time-series data collected when the system is operating normally. While the distribution of healthy samples differs across operating modes within the complex process system, they share semantic consistency and are all in a normal state. Through reversible mapping between healthy sample learning modes, fault samples unique to a particular operating mode can be transferred to other operating modes. The operating modes corresponding to different systems can be set according to the actual scenario. For example, for a nuclear power plant reactor cooling system, operating modes can be categorized into different power load levels, different ambient temperature conditions, and different fuel ratio conditions. Each operating mode corresponds to an independent set of healthy samples, and the sample data can include multi-dimensional monitoring parameters such as temperature, pressure, flow rate, and vibration. It is understood that the collection process of healthy samples in the sample library can cover the complete normal operating cycle, ensuring that the data distribution fully represents the normal operating characteristics of the corresponding mode.

[0050] This embodiment specifically determines the set of healthy samples corresponding to each operating mode from the sample library. For example, if there are k operating modes, the set of healthy samples corresponding to the i-th operating mode is... .

[0051] Step 220: Construct a reversible mapping model for each operating mode based on a reversible neural network, input the set of healthy samples into the corresponding reversible mapping model, and obtain the mapping distribution of the set of healthy samples in the same latent space.

[0052] In one embodiment, the reversible neural network can adopt a flow model architecture with bijective properties. Each operating mode corresponds to an independent reversible mapping model, and the output space of all models is the same shared latent space. The reversible mapping model has both forward and inverse mapping capabilities. Forward mapping transforms the data in the original variable space into feature vectors in the latent space, and inverse mapping restores the feature vectors in the latent space back to the data in the original variable space. The entire mapping process does not result in information loss. For example, K operating modes correspond to K reversible mapping models. The corresponding inverse mapping function is .

[0053] After setting the reversible mapping model corresponding to each operating mode, the parameters of this model are initial parameters, which need to be determined through subsequent training. Here, the healthy sample sets of each operating mode are input into the corresponding reversible mapping model to obtain the feature distribution of each healthy sample set in the shared latent space, and the feature distribution is used as the mapping distribution. In the initial stage of training, the distribution of healthy sample sets in the latent space of different operating modes may differ significantly.

[0054] Step 230: With the goal of minimizing the maximum mean difference between any two mapping distributions, adjust the parameters of all invertible mapping models to obtain the trained invertible mapping model.

[0055] In one embodiment, the training process uses the maximum mean difference as a measure of distribution difference, that is, adjusting the parameters of each invertible mapping model and traversing all combinations of operating modes to minimize the maximum mean difference between any two mapping distributions. The maximum mean difference is used to quantify the distance between two probability distributions in the feature space.

[0056] The specific optimization objective is expressed as follows:

[0057] in, For the i-th operating mode, there is an invertible mapping model. For the j-th operating mode, there is an invertible mapping model. Let i be the set of healthy samples for the i-th operating mode. For the set of healthy samples in the j-th operating mode, This is the square of the largest mean difference.

[0058] In this embodiment, the training process uses the gradient descent algorithm to iteratively update the network parameters of all invertible mapping models until the optimization objective converges. After training, the distribution of healthy samples in all operating modes in the shared latent space basically overlaps, enabling feature alignment of normal states across modes.

[0059] Step 120: Obtain a unique fault sample from one of the operating modes, obtain the initial invertible mapping model corresponding to the initial mode to which the unique fault sample belongs, and calculate the latent space vector corresponding to the unique fault sample in the initial invertible mapping model; obtain the target invertible mapping model corresponding to the target mode, use the inverse mapping function of the target invertible mapping model to restore the latent space vector to the variable space of the target mode, obtain the corresponding target mode fault sample, and update the sample library based on the target mode fault sample.

[0060] In one embodiment, a unique fault sample is a fault sample that is labeled or obtainable only in certain operating modes. For example, a unique fault sample existing only in the i-th operating mode is... It is possible to perform cross-condition feature remapping based on this unique fault sample.

[0061] In one embodiment, an initial invertible mapping model corresponding to the initial mode to which the unique fault sample belongs is obtained, and the latent space vector corresponding to the unique fault sample in the initial invertible mapping model is calculated. The operating mode to which the unique fault sample belongs is called the initial mode, and the operating mode corresponding to the migration target of the unique fault sample is called the target mode. Suppose that it is necessary to generate the corresponding sample in the j-th operating mode from the unique fault sample that exists only in the i-th operating mode through the cross-condition implicit feature remapping mechanism, then the initial mode is the i-th operating mode, and the target mode is the j-th operating mode.

[0062] Next, the reversible mapping model corresponding to the i-th operating mode is loaded as the initial reversible mapping model. The unique fault sample is input into the initial reversible mapping model for forward calculation and deduction to obtain the latent space vector under the unified shared dimension.

[0063] The specific calculation process is as follows:

[0064] Where z is a latent space vector. This represents the initial invertible mapping model corresponding to the i-th operating mode. The latent space vector is the fault semantic feature obtained after completely stripping away the specific working condition background information of the i-th operating mode.

[0065] In one embodiment, a target invertible mapping model corresponding to the target mode is obtained. The inverse mapping function of the target invertible mapping model is used to restore the latent space vector to the variable space of the target mode, thereby obtaining the corresponding target mode fault sample. Specifically, the invertible mapping model corresponding to the j-th operating mode, i.e., the target mode, is obtained as the target invertible mapping model. Since the invertible mapping models of all modes are aligned to the same shared latent space Z, the target invertible mapping model can directly receive the latent space vector z calculated above. Utilizing the mathematical invertibility of the network structure, the inverse mapping function of the target invertible mapping model is called to perform inverse space restoration on the latent space vector z, restoring the latent space vector z to the variable space of the target mode, thereby obtaining the corresponding target mode fault sample.

[0066] The specific calculation process is as follows:

[0067] in, Describes the inverse mapping function of the target invertible mapping model. This represents a sample of the target mode failure.

[0068] As can be seen, after inverse mapping calculation in this embodiment, the latent space vector is decoded and converted into the working condition background features of the target mode, thereby restoring the corresponding data form in the variable space of the target mode. This conversion and restoration mechanism does not rely on the strict assumption that fault categories must completely overlap across different process modes at its underlying logic. In complex industrial environments, even if certain operating modes naturally lack corresponding fault sample data, this conversion mechanism can still be used to achieve cross-condition fault knowledge transfer without requiring complete overlap of fault categories, effectively solving the problem of missing cross-condition fault samples. This expands the scale of fault samples in each process mode, fills data gaps, and eliminates the lack of diagnostic data caused by production environment switching and data distribution offsets at the root, facilitating the construction of high-precision fault diagnosis models.

[0069] In one embodiment, whether to add the target mode fault sample to the sample library requires further verification. (See also...) Figure 3 , Figure 3 This is a flowchart of updating the sample library based on target mode fault samples provided in this application embodiment, specifically including the following steps: Step 310: Obtain any healthy sample in the target mode, and use interpolation coefficients to interpolate the hidden representations corresponding to the healthy sample and the faulty sample in the target mode to obtain the synthetic hidden representation.

[0070] In one embodiment, interpolation is used to simulate the gradual evolution of equipment in a complex process system from a healthy state to a faulty state. First, any healthy sample from the target pattern is obtained. Then, the latent representation of the healthy sample and the latent representation of the faulty sample from the target pattern, generated above, are extracted from the shared latent space of the target pattern. Finally, linear interpolation is performed using interpolation coefficients to obtain the synthesized latent representation.

[0071] The interpolation process is represented as follows:

[0072] in, This is a synthetic implicit representation. This is a hidden representation of healthy samples. This is the implicit representation of the target mode fault sample. These are the interpolation coefficients, which follow a truncated beta distribution with a lower bound. .

[0073] Step 320: Use the inverse mapping model of the target invertible mapping model to restore the synthetic implicit representation to the variable space of the target pattern to obtain the synthetic sample.

[0074] In one embodiment, after the interpolation and fusion of the latent representations in the shared latent space are completed, it is also necessary to convert the synthetic latent representations with abstract semantics into data forms with actual physical meaning. Therefore, the inverse mapping model of the target reversible mapping model corresponding to the target pattern is invoked to perform spatial mapping on the synthetic latent representations, resulting in synthetic samples containing actual parameters such as specific sensor measurements and process operating parameters.

[0075] Step 330: Obtain the process violation degree of the synthetic sample. If the process violation degree meets the preset threshold condition, add the target mode failure sample to the sample library. Otherwise, resample the interpolation coefficients and repeat the test within the preset retry limit until the preset threshold condition is met or the target mode failure sample is abandoned.

[0076] In one embodiment, since simple latent space data interpolation lacks underlying physical constraints, the directly generated synthetic samples may violate objective physical laws. Therefore, this embodiment utilizes a residual network to quantitatively evaluate the process violation degree of the synthetic samples. The process violation degree characterizes the degree of deviation of the synthetic samples from preset physical constraints, which can be material balance errors, energy conservation deviations, excessive temperature change rates, etc. In this embodiment, the process violation degree ψ is defined as the weighted sum of deviations from multiple physical constraints, ψ = Σw k ·e k , where e k This is the normalized deviation value of the k-th physical constraint.

[0077] This embodiment sets a preset threshold condition corresponding to the process violation according to the actual process scenario. This preset threshold condition is expressed as the process violation. It must be less than or equal to the preset tolerance threshold. Preset tolerance threshold The process is set according to the actual scenario. When the preset threshold condition is met, it proves that the physical logic of the currently generated synthetic sample is reasonable, and the target mode fault sample that has not undergone interpolation can be stored in the sample library as a valid sample of the target mode. If the process violation exceeds the preset tolerance threshold, it is determined that the current synthetic sample does not meet the physical constraints, and the target mode fault sample is rejected from being added to the sample library. Furthermore, a retry count is set, for example, a preset retry count of three. Within the maximum retry count limit of three, interpolation coefficients are randomly generated again, a new synthetic sample is constructed according to the above process, and the process violation verification step is repeated. If the preset threshold condition is still not met after reaching the maximum retry count limit, the target mode fault sample generated this time is discarded.

[0078] In one embodiment, reference is made to Figure 4 , Figure 4 This is a schematic diagram of the sample expansion process provided in this embodiment.

[0079] First, latent space interpolation is performed. The latent representations of healthy samples and target mode fault samples are obtained separately. Then, linear interpolation is calculated between the latent representations of healthy samples and target mode fault samples. This linear interpolation fuses the healthy state and the target fault state at the feature level, generating a corresponding synthetic latent representation. Next is inverse mapping and physical constraint verification. The generated synthetic latent representation is received and inversely mapped to the variable space of the target mode using an inverse mapping mechanism, resulting in a synthetic sample with actual physical dimensions and operating condition characteristics. It is then determined whether the current synthetic sample meets the process constraint verification. If the verification result is yes, the sample is determined to conform to physical laws, and the synthetic sample is accepted and added to the sample library. If the verification result is no, the sample is determined to have physical logic deviations, and the sample is rejected, and the interpolation coefficients are retried. When a retry is triggered, it is further determined whether the number of retries is within the preset number of retries. If it is within the preset number of retries, the interpolation coefficients are updated and the process returns along the closed-loop path to perform linear interpolation. If it is determined that the preset number of retries has been exceeded, the sample is discarded, and the generation process terminates.

[0080] As can be seen from the above, compared to the data augmentation and generation methods in related technologies that blindly synthesize data and introduce non-physical noise samples that violate fundamental physical laws, leading to a large number of false alarms in actual industrial operation, this embodiment, by introducing a mechanism of physical verification through process prior constraints after linear interpolation in the latent space, not only ensures that all newly added target mode fault samples in the sample library follow objective physical laws, but also simulates the real gradual physical processes that exist when complex equipment deteriorates.

[0081] Step 130: Use the updated sample library to train the initial fault diagnosis model until a trained fault diagnosis model is obtained.

[0082] In one embodiment, the updated sample library contains healthy samples and fault samples corresponding to each operating mode. The updated sample library is used to train the initial fault diagnosis model until a trained fault diagnosis model is obtained. The fault diagnosis model can perform online fault prediction on the online-acquired data to be detected during the inference phase, thus obtaining a fault diagnosis result.

[0083] The following describes the online real-time diagnostic process of the fault diagnosis model after training.

[0084] In one embodiment, the fault diagnosis model includes at least a dynamic pattern decoupling normalization layer, a multi-scale spatiotemporal feature extraction layer, a dual-stream collaborative focusing layer, and a classification decision layer, wherein the multi-scale spatiotemporal feature extraction layer includes multiple parallel deep convolutional layers with different kernel sizes. (Refer to...) Figure 5 , Figure 5This is a flowchart illustrating how the fault diagnosis model provided in this application performs online fault prediction on online acquired data to be detected, specifically including the following steps: Step 510: Obtain the input features corresponding to the data to be detected, input the input features into the dynamic mode decoupling normalization layer for dynamic normalization, and obtain the normalized output vector.

[0085] In one embodiment, vector extraction is performed on the acquired data to be detected to obtain corresponding input features. These input features are then input into a fault diagnosis model for online real-time diagnosis. The input features first enter a dynamic mode decoupling normalization layer. (Refer to...) Figure 6 , Figure 6 This application provides a flowchart of a process for dynamically normalizing input features into a dynamic pattern decoupling normalization layer to obtain a normalized output vector, which specifically includes the following steps: Step 610: After performing global average pooling on the input features, the input pattern prediction network makes predictions to obtain the posterior probability of the output vector for each operating mode.

[0086] In one embodiment, in complex process systems, the mean and variance of features under different operating modes often exhibit systematic shifts due to feature scarcity. Simply using batch normalization or instance normalization either fails to completely eliminate mode differences or simultaneously eliminates crucial fault-detection features during the difference-elimination process. Therefore, this embodiment, after acquiring the input features, first performs global average pooling along the channel dimension to extract mode-related statistics that reflect the global distribution state of the current input features. The extracted pattern-related statistics are then input into a pre-built pattern prediction network. This network calculates the posterior probability that the input feature belongs to each known operating pattern through forward propagation. Here, the posterior probability represents the confidence level of the pattern prediction network in determining that the current input feature is generated by a specific operating condition.

[0087] Step 620: Obtain the learnable mean and standard deviation parameters corresponding to each operating mode, and perform weighted summation on the mean and standard deviation parameters based on the posterior probability to obtain the weighted mean and weighted standard deviation of the current sample.

[0088] In one embodiment, after obtaining the posterior probabilities of each known operating mode, the predicted posterior probabilities are used as dynamic weighting factors to perform a weighted summation of the learnable mean and standard deviation parameters specific to each operating mode, thereby obtaining the weighted mean and weighted standard deviation of the current sample. The calculation process is expressed as follows:

[0089]

[0090]

[0091] Where K represents the total number of operating modes, and i represents the i-th operating mode. and These represent the learnable mean and standard deviation parameters corresponding to the i-th mode, respectively. They are parameters initially set for the i-th operating mode and continuously updated as the network trains. This represents the posterior probability that the input feature belongs to the i-th operating mode. This represents the weighted average. This represents the weighted standard deviation.

[0092] Step 630: Center and scale the input features using the weighted mean and weighted standard deviation to obtain a normalized output vector.

[0093] In one embodiment, the weighted mean and weighted standard deviation are directly applied to the original input features to center and scale the input features, resulting in a normalized output vector, represented as:

[0094] in, This represents the input features. This represents the normalized output vector. This represents a very small constant used to prevent mathematical overflow errors caused by division by zero.

[0095] In one embodiment, reference is made to Figure 7 , Figure 7 This is a schematic diagram of the processing flow of the dynamic pattern decoupling normalization layer provided in this application embodiment. First, it receives input features acquired in real time and performs global average pooling to extract statistics reflecting the global state distribution of the features. Then, the statistics obtained after global average pooling are input into the pattern prediction network for forward inference, outputting the posterior probability of the input features belonging to each known operating mode. Simultaneously, learnable mean and standard deviation parameters, independently set for each operating mode, are acquired. Using the predicted posterior probability as dynamic weights, a weighted summation operation is performed on the learnable mean and standard deviation parameters. This weighted summation operation achieves adaptive fusion of operating condition statistical parameters, thereby calculating a weighted mean and weighted standard deviation highly adapted to the input features. Next, the input features are combined with the calculated weighted mean to complete the feature centering operation, eliminating systematic mean shifts caused by different process conditions. Finally, the weighted standard deviation is used to further scale the features after the centering operation to smooth the data variance differences of the features. After centering and scaling, a normalized output vector is generated.

[0096] As can be seen from the above, the dynamic mode decoupling normalization layer of this embodiment, without sacrificing the weak fault discrimination capability, removes and suppresses the systematic feature shift between modes through feature adaptive recombination. This allows complex process systems to suppress systematic feature shift without preset mode labels when facing long-term drift or frequent switching of operating conditions, effectively retaining fault discrimination information, and can adaptively process multi-mode data online, thereby improving the cross-operating condition generalization capability of the diagnostic architecture.

[0097] Step 520: Input the normalized output vector into each deep convolutional layer in parallel to extract local temporal features at different scales, and concatenate them in the channel dimension to obtain a multi-scale temporal feature vector.

[0098] In one embodiment, the normalized output vector is input in parallel into multiple deep convolutional layers with different kernel sizes to extract local temporal features at different scales. These features are then concatenated along the channel dimension to obtain a multi-scale temporal feature vector, represented as follows:

[0099] in, Represents multi-scale temporal feature vectors. This represents a depthwise convolution with a kernel size of k. This involves channel splicing. It is understood that the kernel size can be determined based on the actual computational scenario. The above illustrations show four different kernel sizes: 3, 5, 7, and 9, but this does not imply a limitation in this embodiment. This embodiment uses a multi-scale spatiotemporal feature extraction layer to capture fast transient-related features using small kernels and slow drift-related features using large kernels, without significantly increasing computational load.

[0100] Step 530: Input the multi-scale temporal feature vector into the dual-stream co-focusing layer to perform spatiotemporal joint attention calculation to obtain enhanced features.

[0101] In one embodiment, reference is made to Figure 8 , Figure 8 This application provides a flowchart of the process of inputting multi-scale temporal feature vectors into a dual-stream collaborative focusing layer for spatiotemporal joint attention calculation to obtain enhanced features, which specifically includes the following steps: Step 810: Adaptively pool the multi-scale temporal feature vectors along the channel direction to obtain the spatial attention vector, and adaptively pool them along the time direction to obtain the temporal attention vector.

[0102] In one embodiment, a dual-stream collaborative focusing layer is used to focus on key spatiotemporal regions and uncover the inherent correlations within the data. (See also...) Figure 9 , Figure 9This is a schematic diagram of the computational processing flow for enhanced features provided in an embodiment of this application. First, the received multi-scale temporal feature vector is a two-dimensional feature matrix containing channel and time dimensions. This multi-scale temporal feature vector is then fed in parallel into two parallel processing branches: the temporal stream and the spatial stream.

[0103] In the temporal stream branch, multi-scale temporal feature vectors Adaptive pooling is performed along the channel direction to compress the feature map into a single-channel time series, and a temporal attention vector reflecting time fluctuations is extracted. The adaptive pooling process first performs channel-average pooling, then sequentially inputs the data into two fully connected layers to obtain the temporal attention vector. Similarly, in the spatial flow branch, multi-scale temporal feature vectors are processed... Adaptive pooling is performed along the time direction to extract the spatial attention vector that reflects spatial dependencies. Here, adaptive pooling processing is consistent with the temporal stream branch. This embodiment, through this parallel dual-stream architecture, can capture key feature information from both the temporal and spatial dimensions.

[0104] Step 820: Obtain the spatiotemporal joint attention vector based on the temporal attention vector and the spatial attention vector. Multiply the spatiotemporal joint attention vector and the multi-scale temporal feature vector element by element and sum them according to the time dimension to obtain the enhanced features.

[0105] In one embodiment, reference is made to Figure 9 By using matrix outer product, temporal and spatial attention vectors are fused, preserving complete spatiotemporal coupling information and providing targeted enhancement to multivariate collaborative anomaly regions in the enhanced features. Specifically, the outer product of the transposes of the spatial and temporal attention vectors is calculated to generate a two-dimensional spatiotemporal joint attention vector. , represented as:

[0106] After generating the spatiotemporal joint attention vector, it is multiplied element-wise with the input multi-scale temporal feature vector using a Hadamard product. This leverages the weight distribution of the spatiotemporal joint attention map to target and enhance key anomalous regions of the multi-scale temporal feature vector while suppressing irrelevant background noise. Finally, the multiplied feature matrix is ​​summed and dimensionality reduced along the time dimension to output the final enhanced feature.

[0107] As can be seen, this embodiment utilizes a dual-stream structure to retain complete spatiotemporal coupling information. The joint attention map generated through matrix outer product can capture the abnormal coupling state between different sensors in the spatial dimension and the abnormal evolution trend in the temporal dimension. It is suitable for detecting complex fault modes that require joint judgment of multiple variables, improving the fault identification accuracy of complex process systems under multi-variable collaborative anomaly conditions.

[0108] Step 540: Use the classification decision layer to predict the probability of fault categories on the enhanced features to obtain the fault diagnosis results.

[0109] In one embodiment, the final enhanced features are used to predict the fault category probability through a classification decision layer consisting of a fully connected layer and a softmax layer, outputting the corresponding fault diagnosis result. The fault diagnosis result is obtained based on the maximum fault category probability, specifically using the following formula:

[0110]

[0111]

[0112] Among them, fault diagnosis results It is obtained based on the maximum probability of the fault category and provides the confidence level. . This represents the probability distribution vector of the fault categories calculated by the classification decision layer. Each component in this vector directly corresponds to the probability of occurrence of a specific fault category. This represents the learnable weight matrix of a fully connected layer network. Represents the augmentation vector. This represents the learnable bias vector of a fully connected layer network. This represents the physical fault label corresponding to the dimension index position of the probability component with the largest retrieved value in the probability distribution vector, which is the fault diagnosis result. Additionally, the maximum value in the probability distribution vector is selected as the confidence level. .

[0113] The above process describes the online real-time diagnosis of real-time data by the fault diagnosis model trained in this embodiment. To enable the fault diagnosis model to adapt to slow process changes and maintain high accuracy over a long period, this embodiment also performs adaptive updates and fine-tuning. (Refer to...) Figure 10 , Figure 10 This is a flowchart of the adaptive update of the fault diagnosis model provided in the embodiments of this application, which specifically includes the following steps: Step 1010: Obtain the confidence level of the fault diagnosis result output by the fault diagnosis model after performing online fault prediction on the data to be detected. If the confidence level is greater than or equal to the preset confidence level threshold, combine the data to be detected and the fault diagnosis result into a new sample with pseudo-labels and store it in the cache pool.

[0114] In one embodiment, considering the long-term drift phenomenon in actual industrial processes, such as catalyst aging and deactivation after prolonged operation or scaling on heat exchanger surfaces, these gradual physical changes cause the underlying data distribution to slowly deviate from its initial state, leading to a gradual decrease in the detection accuracy of fault diagnosis models trained entirely on historical data. To combat this long-term drift, this embodiment introduces an online unsupervised and semi-supervised fine-tuning mechanism based on high-confidence pseudo-labels.

[0115] Specifically, the fault diagnosis results and corresponding confidence levels output during the online real-time diagnosis process are obtained, and the confidence level c is compared with a pre-set confidence threshold. The comparison is performed, and the confidence threshold is used here. It can be set between 0.9 and 0.98, for example, a specific value of 0.95. If the comparison result shows... If the fault diagnosis result of the data to be detected is deemed to be relatively reliable, the highly reliable fault diagnosis result is directly extracted into a pseudo-label and bound to the original data to be detected to form a new sample with supervision information. Finally, the new sample is uniformly sent to a specially allocated cache pool for temporary storage and accumulation.

[0116] Step 1020: When the number of new samples accumulated in the cache pool reaches the preset threshold, the sum of cross-entropy loss and regularization loss is used as the total loss function. The new samples in the cache pool are used to fine-tune the fault diagnosis model in a small batch. After the fine-tuning is completed, the cache pool is cleared.

[0117] In one embodiment, the number M of new samples accumulated in the cache pool is monitored in real time. When the number of new samples in the cache pool reaches a preset number to trigger the fine-tuning mechanism, for example, 500, the online update process of the model parameters is initiated after accumulating 500 new samples. During the update optimization process, the sum of cross-entropy loss and an additional regularization loss is used as the total loss function for overall optimization. Here, cross-entropy loss is introduced to enable the fault diagnosis model to quickly fit and learn the pseudo-label distribution features of new samples in the cache pool. The regularization loss is introduced to effectively prevent catastrophic forgetting in the process of learning new data features, avoiding forgetting important historical fault knowledge accumulated during the offline training phase. Therefore, the regularization loss is used to constrain the parameter update amplitude of the fault diagnosis model to prevent catastrophic forgetting. The regularization loss can be quantified by the divergence value between the output probability distribution of the model parameters before and after the update on the validation set, expressed as:

[0118] in, This represents the probability distribution of the new model's output for a given input feature after parameter updates. This represents the initial probability distribution state of the old model's output for a given input feature before the parameter update.

[0119] Next, the total loss function is used to perform a mini-batch fine-tuning operation on the fault diagnosis model. After all the mini-batch fine-tuning operations are completed and the network weight updates converge, all temporary data in the current cache pool is completely cleared, and the next round of real-time sample collection and adaptive update loop continues.

[0120] As can be seen, this application's embodiments, through a cross-operating condition feature remapping mechanism, can remap unique fault samples to target mode fault samples using only healthy samples from each operating mode, effectively solving the problem of detection failure due to missing fault samples and improving the multi-mode data coverage of the sample library. Furthermore, based on a physical interpolation synthesis strategy, after generating and inversely mapping synthetic samples, physical verification is performed using process violation to ensure that the samples in the library conform to physical gradual change laws, improving the model's sensitivity to early, subtle faults. It also utilizes a dynamic mode decoupling normalization layer to adaptively calculate weighted statistical parameters, effectively suppressing operating condition offset interference. Combined with multi-scale spatiotemporal feature extraction and a dual-stream collaborative focusing layer, it captures temporal evolution features and focuses on spatiotemporally joint anomaly regions. Simultaneously, it generates data with pseudo-labels by evaluating confidence levels and performs fine-tuning training using cross-entropy loss and regularization loss. While using online data to be detected to evolve model parameters, it effectively prevents the forgetting of historical fault knowledge, ensuring high accuracy of the fault diagnosis model throughout its entire lifecycle.

[0121] In one embodiment, reference is made to Figure 11 , Figure 11 This is an overall flowchart of the multi-mode adaptive fault diagnosis provided in the embodiments of this application. The overall process can be divided into an offline cross-operating condition knowledge transfer stage, an online real-time diagnosis stage, and a model adaptive update stage.

[0122] In the offline cross-operating condition knowledge transfer phase, historical data for various operating modes in the sample library are acquired and divided into healthy samples and unique fault samples. Reversible mapping models are trained based on the corresponding healthy samples, ensuring that all reversible mapping models map healthy samples from different operating modes to the same shared latent space. Then, at least two corresponding reversible mapping models are used to perform cross-operating condition feature remapping on unique fault samples under specific modes, generating target mode fault samples and transferring fault features under unknown operating conditions across domains. Simultaneously, external physical constraints are used to interpolate and synthesize the target mode fault samples for verification, and the sample library is updated and expanded based on the verification results.

[0123] During the online real-time diagnosis phase, real-time acquired data is used as the data to be detected and input into the trained fault diagnosis model for inference. The data to be detected sequentially enters the dynamic mode decoupling normalization layer, the multi-scale spatiotemporal feature extraction layer, and the dual-stream collaborative focusing layer for refinement and spatiotemporal focusing to obtain enhanced features. The enhanced features enter the classification decision layer to output the fault category of the current data to be detected as the fault diagnosis result, and simultaneously outputs a confidence level for quantitatively evaluating the fault diagnosis result.

[0124] During the model adaptive update phase, the confidence level transmitted from the online real-time diagnosis phase is continuously received and monitored. When the confidence level corresponding to the current diagnosis result is determined to be greater than or equal to a preset confidence threshold, pseudo-labels with supervisory guidance significance are generated based on the data to be detected and the diagnosis result. These accumulated pseudo-labeled data are added to the cache pool. When the accumulation reaches a preset number, this data is used as driving data. The fault diagnosis model is adaptively fine-tuned and updated using a total loss function that includes cross-entropy loss and a regularization term to prevent forgetting, along with the pseudo-labeled data. Finally, the cache pool is cleared. This three-stage collaborative operation realizes a complete process from offline knowledge transfer to online diagnosis and then to continuous adaptation.

[0125] Furthermore, advanced persistent threat attacks targeting critical infrastructure such as nuclear power plants often exploit vulnerabilities in industrial control systems to tamper with sensor data or inject false control commands, leading to abnormal process parameters. The behavioral anomalies caused by such attacks are highly similar to failure modes. Therefore, the multi-mode adaptive fault diagnosis method in this embodiment can also be used for process behavior perception within a network security monitoring system, filling the gap where security monitoring only focuses on network traffic and lacks awareness of physical process anomalies.

[0126] The technical solution provided in this application involves determining a reversible mapping model for each operating mode based on the corresponding healthy samples in the sample library, with all reversible mapping models sharing the same latent space. A unique fault sample for one operating mode is obtained, along with the initial reversible mapping model corresponding to the initial mode to which the unique fault sample belongs. The latent space vector corresponding to the unique fault sample in the initial reversible mapping model is calculated. A target reversible mapping model for the target mode is obtained, and the latent space vector is restored to the variable space of the target mode using the inverse mapping function of the target reversible mapping model to obtain the corresponding target mode fault sample. The sample library is updated based on the target mode fault sample. The updated sample library is used to train the initial fault diagnosis model until a trained fault diagnosis model is obtained. The fault diagnosis model is used to perform online fault prediction on the online acquired data to be detected, obtaining the fault diagnosis result. This application embodiment constructs reversible mapping models sharing the same latent space using healthy samples corresponding to each operating mode, achieving distribution alignment of different modes in the latent space using only healthy samples. Based on this, unique fault samples from a single mode are transferred to other target modes through cross-condition feature remapping, generating corresponding fault samples for the target modes and expanding the sample library to fill the gaps in the corresponding fault samples for the target modes. The fault diagnosis model trained based on the expanded sample library has good cross-condition generalization ability, can adapt to mode switching scenarios under different loads and operating conditions, and improves the fault diagnosis accuracy of multi-mode complex process systems.

[0127] This application also provides a multi-mode adaptive fault diagnosis device that can implement the above-described multi-mode adaptive fault diagnosis method, see reference. Figure 12 The device includes: Reversible mapping module 1210: For each operating mode, a reversible mapping model is determined based on the corresponding healthy sample in the sample library. All reversible mapping models share the same latent space.

[0128] Feature remapping module 1220: used to obtain unique fault samples of one of the operating modes, perform cross-operating condition feature remapping on the unique fault samples according to at least two corresponding reversible mapping models to obtain target mode fault samples, and update the sample library based on the target mode fault samples.

[0129] Fault Diagnosis Training Module 1230: This module is used to train the initial fault diagnosis model using the updated sample library until a well-trained fault diagnosis model is obtained. The fault diagnosis model is used to perform online fault prediction on the online acquired data to be detected, and obtain the fault diagnosis result.

[0130] The reversible mapping module is used to transfer the parameters of each reversible mapping model after training to the feature remapping module; the feature remapping module is used to transfer the updated sample library to the fault diagnosis training module; the fault diagnosis training module is used to deploy the trained fault diagnosis model for online prediction and receive new samples that meet the confidence conditions during the online diagnosis process to incrementally fine-tune the fault diagnosis model. The fine-tuned parameters are fed back to the online prediction stage to form a closed loop.

[0131] The specific implementation of the multi-mode adaptive fault diagnosis device in this embodiment is basically the same as the specific implementation of the multi-mode adaptive fault diagnosis method described above, and will not be repeated here.

[0132] This application also provides an electronic device, including: At least one memory; At least one processor; At least one program; The program is stored in a memory, and the processor executes the at least one program to implement the multi-mode adaptive fault diagnosis method described above in this application. The electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0133] Please see Figure 13 , Figure 13 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 1301 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 1302 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1302 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1302 and is called and executed by the processor 1301 using the multi-mode adaptive fault diagnosis method of the embodiments of this application. The input / output interface 1303 is used to implement information input and output; The communication interface 1304 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 1305 transmits information between various components of the device (e.g., processor 1301, memory 1302, input / output interface 1303, and communication interface 1304); The processor 1301, memory 1302, input / output interface 1303 and communication interface 1304 are connected to each other within the device via bus 1305.

[0134] This application also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described multi-mode adaptive fault diagnosis method.

[0135] Memory, as a non-transitory storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0136] The multi-mode adaptive fault diagnosis method, apparatus, device, and storage medium proposed in this application, for each operating mode, determines a reversible mapping model based on the corresponding healthy samples in the sample library, with all reversible mapping models sharing the same latent space; obtains a unique fault sample for one operating mode, obtains the initial reversible mapping model corresponding to the initial mode to which the unique fault sample belongs, and calculates the latent space vector corresponding to the unique fault sample in the initial reversible mapping model; obtains the target reversible mapping model corresponding to the target mode, and uses the inverse mapping function of the target reversible mapping model to restore the latent space vector to the variable space of the target mode to obtain the corresponding target mode fault sample; updates the sample library based on the target mode fault sample; and trains the initial fault diagnosis model using the updated sample library until a trained fault diagnosis model is obtained. The fault diagnosis model is used to perform online fault prediction on the online acquired data to be detected, and obtains the fault diagnosis result. This application embodiment constructs reversible mapping models sharing the same latent space using healthy samples corresponding to each operating mode, and the distribution alignment of different modes in the latent space can be completed using only healthy samples. Based on this, unique fault samples from a single mode are transferred to other target modes through cross-condition feature remapping, generating corresponding fault samples for the target modes and expanding the sample library to fill the gaps in the corresponding fault samples for the target modes. The fault diagnosis model trained based on the expanded sample library has good cross-condition generalization ability, can adapt to mode switching scenarios under different loads and operating conditions, and improves the fault diagnosis accuracy of multi-mode complex process systems.

[0137] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0138] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0139] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0140] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0141] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0142] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0143] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0144] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0145] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0146] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0147] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A multi-mode adaptive fault diagnosis method, characterized in that, include: For each operating mode, a reversible mapping model is determined based on the corresponding healthy samples in the sample library, and all the reversible mapping models share the same latent space; Obtain a unique fault sample of one of the operating modes, obtain the initial reversible mapping model corresponding to the initial mode to which the unique fault sample belongs, and calculate the latent space vector corresponding to the unique fault sample in the initial reversible mapping model; obtain the target reversible mapping model corresponding to the target mode, and use the inverse mapping function of the target reversible mapping model to restore the latent space vector to the variable space of the target mode to obtain the corresponding target mode fault sample; update the sample library based on the target mode fault sample. The initial fault diagnosis model is trained using the updated sample library until a trained fault diagnosis model is obtained. The fault diagnosis model is used to perform online fault prediction on the online acquired data to be detected, and obtain fault diagnosis results.

2. The multi-mode adaptive fault diagnosis method according to claim 1, characterized in that, The step of determining the reversible mapping model based on the corresponding healthy samples in the sample library includes: Determine the set of healthy samples corresponding to each of the operating modes from the sample library; Based on the reversible neural network, a reversible mapping model is constructed for each of the operating modes. The set of healthy samples is input into the corresponding reversible mapping model to obtain the mapping distribution of the set of healthy samples in the same latent space. With the objective of minimizing the maximum mean difference between any two of the said mapping distributions, the parameters of all the said invertible mapping models are adjusted to obtain the trained invertible mapping model.

3. The multi-mode adaptive fault diagnosis method according to claim 1, characterized in that, The updating of the sample library based on the target mode fault samples includes: Obtain any one of the healthy samples in the target mode, and use interpolation coefficients to interpolate the hidden representations corresponding to the healthy sample and the faulty sample of the target mode to obtain a synthetic hidden representation; The synthetic implicit representation is restored to the variable space of the target pattern using the inverse mapping model of the target invertible mapping model to obtain the synthetic sample; The process violation degree of the synthesized sample is obtained. If the process violation degree meets the preset threshold condition, the target mode failure sample is added to the sample library; otherwise, the interpolation coefficient is resampled and the test is repeated within the retry limit until the preset threshold condition is met or the target mode failure sample is abandoned.

4. The multi-mode adaptive fault diagnosis method according to claim 1, characterized in that, The fault diagnosis model includes at least a dynamic pattern decoupling normalization layer, a multi-scale spatiotemporal feature extraction layer, a dual-stream collaborative focusing layer, and a classification decision layer. The multi-scale spatiotemporal feature extraction layer includes multiple parallel deep convolutional layers with different kernel sizes. The fault diagnosis model is used to perform the following steps when performing online fault prediction on online acquired data to be detected: The input features corresponding to the data to be detected are obtained, and the input features are input into the dynamic mode decoupling normalization layer for dynamic normalization to obtain a normalized output vector. The normalized output vector is input in parallel into each of the deep convolutional layers to extract local temporal features at different scales, and then concatenated along the channel dimension to obtain a multi-scale temporal feature vector. The multi-scale temporal feature vector is input into the dual-stream collaborative focusing layer for spatiotemporal joint attention calculation to obtain enhanced features; The classification decision layer is used to predict the probability of fault categories based on the enhanced features, thereby obtaining fault diagnosis results.

5. The multi-mode adaptive fault diagnosis method according to claim 4, characterized in that, The step of inputting the input features into the dynamic pattern decoupling normalization layer for dynamic normalization to obtain a normalized output vector includes: After performing global average pooling on the input features, the input pattern prediction network makes predictions to obtain the posterior probability of the output vector for each of the operating modes. Obtain the learnable mean and standard deviation parameters corresponding to each of the operating modes, and perform a weighted summation of the mean and standard deviation parameters based on the posterior probability to obtain the weighted mean and weighted standard deviation of the current sample. The input features are centered and scaled using the weighted mean and the weighted standard deviation to obtain the normalized output vector.

6. The multi-mode adaptive fault diagnosis method according to claim 4, characterized in that, The step of inputting the multi-scale temporal feature vector into the dual-stream collaborative focusing layer for spatiotemporal joint attention calculation to obtain enhanced features includes: The multi-scale temporal feature vectors are adaptively pooled along the channel direction to obtain spatial attention vectors, and adaptively pooled along the time direction to obtain temporal attention vectors. The spatiotemporal joint attention vector is obtained based on the temporal attention vector and the spatial attention vector. The spatiotemporal joint attention vector and the multi-scale temporal feature vector are multiplied element-wise and summed according to the time dimension to obtain the enhanced feature.

7. The multi-mode adaptive fault diagnosis method according to claim 1, characterized in that, The method further includes: Obtain the confidence level of the fault diagnosis result output by the fault diagnosis model after performing online fault prediction on the data to be detected; If the confidence level is greater than or equal to the preset confidence level threshold, the data to be detected and the fault diagnosis result are combined into a new sample with a pseudo-label and stored in the cache pool. The fault diagnosis model is then fine-tuned in a small batch using the new sample in the cache pool.

8. The multi-mode adaptive fault diagnosis method according to claim 7, characterized in that, The step of using the new samples in the cache pool to perform small-batch fine-tuning of the fault diagnosis model includes: When the number of new samples accumulated in the cache pool reaches a preset threshold, the total loss function is the sum of cross-entropy loss and regularization loss. The fault diagnosis model is then fine-tuned using the new samples in the cache pool in a small batch. After the fine-tuning is completed, the cache pool is cleared.

9. A multi-mode adaptive fault diagnosis device, characterized in that, include: Reversible mapping module: used to determine a reversible mapping model for each operating mode based on the corresponding healthy samples in the sample library, and all the reversible mapping models share the same latent space; Feature remapping module: used to obtain a unique fault sample of one of the operating modes, obtain the initial reversible mapping model corresponding to the initial mode to which the unique fault sample belongs, calculate the latent space vector corresponding to the unique fault sample in the initial reversible mapping model; obtain the target reversible mapping model corresponding to the target mode, use the inverse mapping function of the target reversible mapping model to restore the latent space vector to the variable space of the target mode, obtain the corresponding target mode fault sample, and update the sample library based on the target mode fault sample; Fault diagnosis training module: used to train the initial fault diagnosis model using the updated sample library until a trained fault diagnosis model is obtained. The fault diagnosis model is used to perform online fault prediction on the online acquired data to be detected, and obtain fault diagnosis results.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the multi-mode adaptive fault diagnosis method according to any one of claims 1 to 8.

11. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the multi-mode adaptive fault diagnosis method according to any one of claims 1 to 8.

Images