A method and system for plant equipment failure prediction that fuses causal inference

Abstract

The application relates to a kind of factory equipment failure prediction method and system fusing causal inference, method includes: obtaining local time series data, and input trained local model, generate preliminary causal diagram;Initialize the parameters of global model;Based on preliminary causal diagram, by each client to local factory sensor time series data carries out causal perception data enhancement;Based on the local factory sensor time series data after data enhancement, comparative learning of local model is carried out, and the updated model parameter is uploaded to the server after encryption;By server running knowledge-guided aggregation algorithm, update the global model, and the updated global model is issued to each client;By each client, real-time acquisition local factory sensor time series data under different data sources, input global model, generate universality feature, and send into prediction model, generate fault analysis result.Compared with prior art, the application has the advantages of considering data security and flexible system deployment.

Description

Technical Field

[0001] This invention relates to the field of data analysis, and in particular to a method and system for predicting factory equipment failures by incorporating causal inference. Background Technology

[0002] With the widespread application of big data and artificial intelligence in recommender systems, smart transportation, and the Industrial Internet of Things, data-driven intelligent decision-making systems face increasingly severe challenges. First, large-scale data often contains biases and confounding factors. For example, in recommender systems, the popularity of items and users' historical behavior, as confounding variables, can lead to models learning spurious correlations and non-genuine causal relationships. Similarly, in the medical field, differences between different hospitals or groups (such as age, lifestyle habits, and other implicit factors) can obscure the true drivers of disease risk; in the industrial field, sensor errors or changes in the operating environment can also introduce interference. Second, deep learning models often lack interpretability; their predictions are like "black boxes," making them difficult for decision-makers to understand and trust. In high-risk scenarios such as medical diagnosis or industrial fault prediction, models lacking clear causal logic are often considered unacceptable. Third, data silos and privacy restrictions exacerbate the difficulty of modeling. Data in fields such as healthcare and industry is often scattered across different institutions and is strictly protected by regulations and privacy, making it difficult for traditional centralized learning models to acquire complete data; at the same time, the heterogeneity of data distribution leads to insufficient generalization ability of models when applied across domains.

[0003] Existing technologies each have their shortcomings in addressing the aforementioned problems: purely causal inference methods often suffer from poor scalability in high-dimensional heterogeneous data, making them difficult to apply directly to complex scenarios; while purely contrastive learning methods can extract robust representations, the learned features may not necessarily have causal significance and are prone to capturing non-stationary related features. Furthermore, conventional federated learning lacks utilization of causal structures, making it difficult to guarantee consistent, causally interpretable features when aggregating models across clients. In conclusion, constructing a data engineering method that combines causal inference and contrastive learning, operating collaboratively within a federated learning framework, represents a cutting-edge technological path to improve model reliability, interpretability, and privacy friendliness.

[0004] The invention disclosed in CN119249258A presents a training method, identification method, and apparatus for a sudden event recognition model. It pre-trains a global classifier based on a federated learning architecture, extracts and aggregates local structural information from the local graphs of each client using a random block model, and reconstructs a local view with global information. Through knowledge distillation, the local view with global information guides the model optimization training from a local perspective, minimizing the representational differences between the two graph structures. A contrastive learning method is used to construct an enhanced view by perturbing and expanding the graph data. Representations of corresponding nodes in the original graph and the enhanced view are selected as positive samples, and representations of different nodes are selected as negative samples, constraining the representational distance between positive and negative samples to improve the model's representational learning ability and robustness. However, the predicted information output by this scheme has weak causal correlation, resulting in poor reliability of the generated decisions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art by providing a method and system for predicting equipment failures, diseases, and traffic flow by incorporating causal inference.

[0006] The objective of this invention can be achieved through the following technical solutions: A factory equipment failure prediction method integrating causal inference includes: Step 1: Each client obtains local time-series data from different data sources, inputs it into the trained local model, and generates a preliminary causal graph representing local device faults; the server initializes a pre-built global model for extracting universal features from the input data; the local time-series data is local factory sensor time-series data. Step 2: Based on the preliminary causal graph, each client performs causal perception data augmentation on the local factory sensor time series data; then, each client performs comparative learning of the local model based on the data-augmented local factory sensor time series data to obtain the local causal feature representation, updates the local model parameters based on the causal feature representation, and encrypts and uploads the updated model parameters to the server. Step 3: Based on the updated model parameters, the server runs a knowledge-guided aggregation algorithm to filter and aggregate features corresponding to stable causal relationships, update the parameters of the global model, and distribute the updated global model to each client. Step 4: Each client collects local factory sensor time-series data from different data sources in real time, preprocesses it, and inputs it into the global model to generate feature vectors and perform feature mapping transformation to obtain general features characterizing local equipment faults; the general features are then sent into a pre-built prediction model to generate fault analysis results.

[0007] Furthermore, each client uses the BFS causal discovery method to train the local model, then inputs the local factory sensor time series data, and the trained local model outputs a preliminary causal graph characterizing local equipment faults.

[0008] Furthermore, the server runs a knowledge-guided aggregation algorithm to analyze the consistency of causal structures contained in parameters from different clients, filter and aggregate features corresponding to stable causal relationships in the factory clients, and update the global model.

[0009] Furthermore, the residual adaptive parameter aggregation method is used to aggregate the features corresponding to the causal associations; and when aggregating the features corresponding to the causal associations of each client, the feature representations that are shared by each client and are causally consistent are selected first.

[0010] Furthermore, step three also includes: After the updated global model is distributed to each client, new local factory sensor time-series data is input into the global model. The global model generates data features that are universally representative of local equipment faults, and subsequent fault prediction and path identification are performed based on the data features.

[0011] Furthermore, the importance of the data features is evaluated based on a composite scoring mechanism that combines causal contribution and contrast discrimination, and the final data features with universality are selected. The importance of the data features is mainly determined by their contribution to the final causal effect and their stability under different contrast views.

[0012] Furthermore, step three, the parameter update process for each client, includes: After the updated global model is distributed to each client, fault prediction and path identification data corresponding to the newly collected local factory sensor time series data are collected, and it is determined whether the local model parameters need to be updated. If the pre-set fault prediction and path identification indicators are met, training is not required. If the fault prediction and path identification indicators are not met, the process returns to step two.

[0013] Furthermore, data enhancement for causal perception specifically includes: The data augmentation is guided by the parent nodes in the preliminary causal graph; the local model identifies and generates a key invariant subgraph, and then performs data augmentation based on the key invariant subgraph to generate diverse and causally related positive samples and counterfactual negative samples, which are then used for comparative learning of the local model.

[0014] Furthermore, the negative samples are virtual counterfactual sample data without causal correlation generated based on the key invariant subgraph, so that the local model, after training, can generate causal discriminative features that truly cause equipment failure.

[0015] The present invention also provides a system for a factory equipment failure prediction method that integrates causal inference as described above, comprising: Multiple clients, each acquiring local time-series data from different data sources, inputting it into a trained local model, and generating a preliminary causal graph representing the causal relationships of local device failures; The server is used to initialize the parameters of a pre-built global model for generating generalized features; the local time-series data is local factory sensor time-series data. The local causal discovery module, located on each client, is used to perform causal perception data augmentation on the local factory sensor time series data based on the preliminary causal graph. Then, each client performs comparative learning of the local model based on the data-augmented local factory sensor time series data to obtain the local causal feature representation, updates the local model parameters based on the causal feature representation, and encrypts and uploads the updated model parameters to the server. The knowledge-guided aggregation module, located on the server side, is used to update the global model by running a knowledge-guided aggregation algorithm based on the updated model parameters, and then distribute the updated global model to each client. The prediction module is used to acquire local factory sensor time-series data collected in real time from different data sources by various clients, input the data into the global model, generate universal features characterizing the causal relationship of local equipment failures, and send them into the pre-built prediction model to generate equipment failure analysis results.

[0016] Compared with the prior art, the present invention has the following advantages: (1) This invention constructs a fusion causal inference model for predicting factory equipment failures: First, the model undergoes preliminary training for feature recognition to obtain a preliminary causal graph. Then, causal perception data augmentation is performed on local sensor time-series data. After data augmentation, based on the preliminary causal graph, comparative learning of the local model is conducted to obtain local causal feature representations. Next, causal invariant feature aggregation for federated learning is performed: each client updates its local model parameters, encrypts the updated model parameters, and uploads them to the server. Finally, the server distributes the updated global model to each client. From data preprocessing to model training, the synergistic effect of three core technologies—causal guidance, contrastive learning, and federated aggregation—enhanced the system's performance and interpretability. By introducing causal structures into feature generation and selection, potential confounding factors on the user and item sides were identified and controlled, achieving debiasing at the data representation level. Furthermore, by utilizing the invariant learning concept in causal inference, contrastive learning was guided to focus on core features that are stable across environments and domains, improving the model's robustness in unknown scenarios.

[0017] (2) This invention utilizes prior knowledge provided by causal structure to guide data augmentation and feature selection, performs comparative learning on the model to extract cross-domain invariant representations, and integrates multi-source data through federated learning mechanism, thereby achieving debiasing and generalization in high-dimensional heterogeneous environments. This not only significantly improves the accuracy and robustness of decision-making, but also takes into account the needs of data security and deployment flexibility.

[0018] (3) This invention can play a role in a variety of fields: in the medical field, it can help doctors identify the real key biomarkers of diseases; in the industrial field, it can clearly show the propagation path of faults and help precise operation and maintenance; in the recommendation system, it can effectively alleviate the "Matthew effect", allowing long-tail products and niche content to have a fairer exposure opportunity and improve the health of the ecosystem; when the data distribution changes (such as the recommendation system encountering new users or traffic prediction encountering sudden road conditions), the model can show stronger adaptability and stability by relying on the causal invariance feature; thanks to the natural combination with federated learning, this framework can build joint models among multiple institutions without aggregating sensitive data, accelerating the research process in fields such as drug efficacy evaluation. Attached Figure Description

[0019] Figure 1 This is a flowchart of a factory equipment failure prediction method that integrates causal inference, provided in an embodiment of the present invention. Figure 2 This is an overall flowchart of a factory equipment failure prediction method that integrates causal inference provided in an embodiment of the present invention; Figure 3 This is a simplified flowchart of a factory equipment failure prediction method that integrates causal inference, provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] Definitions: Residual adaptive parameter aggregation is an advanced aggregation method for dynamically adjusting model fusion strategies in federated learning or distributed optimization. Its core idea is to intelligently allocate aggregation weights by analyzing the residuals (differences) between client-uploaded parameters and the global model. This method first calculates the residual vectors between each local model update and the global model. Then, it adaptively adjusts the aggregation coefficients based on the norm, directional consistency, and historical performance of the residuals—giving higher weights to clients with stable residuals and directions consistent with the global optimization, and reducing or eliminating clients with abnormal residuals (potentially due to low-quality data or malicious attacks). This technique effectively resists model drift caused by data heterogeneity, improves convergence stability, and, in privacy-preserving computing scenarios, collaborates with differential noise mechanisms to achieve a more refined privacy-utility tradeoff.

[0024] Knowledge-guided aggregation algorithms are advanced aggregation mechanisms specifically designed within federated learning frameworks for intelligently fusing local knowledge from multiple distributed clients. Instead of simply averaging model parameters, this algorithm first analyzes structured knowledge, such as causal graph structures, learned by each client. Through techniques like graph matching, causal path alignment, and consensus discovery, it identifies stable causal patterns and conflicts across clients. Then, guided by this consensus-based causal knowledge, it selectively and differentially aggregates neural network parameters—for example, assigning higher weights to model parameters that conform to the global causal structure, preserving some autonomy for knowledge based on unique local causal discoveries, or strengthening the fusion of feature dimensions with high causal contributions. Its ultimate goal is to produce a global model that not only has higher accuracy but also stronger interpretability, better generalization ability, and reflects the unified causal mechanism behind the data.

[0025] The Matthew effect: In federated learning, this specifically refers to the polarization of model performance: clients with high-quality data and strong computing power gain stronger representation capabilities in the global model by continuously contributing high-quality gradients, thus performing better in the next round of training, forming a positive feedback loop of "the strong getting stronger"; while clients with poor data or weak computing power are further marginalized due to the aggravation of model bias.

[0026] Basis Function Score: This metric is used to evaluate or quantify the contribution of model components (basis functions) to a target task. In function approximation or model interpretation, it calculates the importance score by analyzing the weights or sensitivities of each basis function (such as polynomial terms, Fourier bases, and neural network neurons) in the output. In federated learning or ensemble systems, this score can be used for adaptive model aggregation, such as assigning fusion weights that match the contribution of local models (considered as basis functions) from different clients, thereby optimizing global model performance and mitigating heterogeneity bias. Its core advantage lies in its ability to achieve more refined contribution attribution and resource allocation through structured model decomposition.

[0027] Paillier is an additive homomorphic public-key cryptosystem based on the compound residues problem. It allows direct addition of ciphertext without decryption and supports multiplication by arbitrary constants. This algorithm is widely used in scenarios such as federated learning and secure multi-party computation, enabling servers to securely aggregate client-encrypted model parameters without decrypting individual uploaded data, thereby ensuring both data privacy and aggregation correctness.

[0028] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

[0029] Example 1 like Figure 1 As shown, this embodiment provides a factory equipment failure prediction method that integrates causal inference. The method includes the following steps: S1: Each client obtains local time-series data from different data sources, inputs it into the trained local model, and generates a preliminary causal graph representing local device faults; the server initializes a pre-built global model for extracting universal features from the input data; the local time-series data is local factory sensor time-series data. S101: Initialization and Data Preparation: Connects clients to multiple industrial sites, each client locally storing sensor time-series data of its device operation. The central server initializes global model parameters; and the input data consists of sensor time-series data from different factories (clients) with similar models but different operating environments.

[0030] S102: Each client uses scalable causal discovery methods such as Basis Function Score to learn a preliminary causal graph from local time series data.

[0031] S2: Based on the initial causal graph, each client performs causal perception data augmentation on the local factory sensor time series data; then, each client performs comparative learning of the local model based on the data-augmented local factory sensor time series data to obtain the local causal feature representation, updates the local model parameters based on the causal feature representation, and encrypts and uploads the updated model parameters to the server. Preferred, Encryption methods can generally employ direct ciphertext computation based on homomorphic encryption (such as the Paillier scheme which supports secure aggregation of encrypted parameters by the server) and distributed collaborative computation based on secure multi-party computation (such as using secret sharing to transmit parameters in fragments, where the server needs to work with multiple fragments to recover the data).

[0032] S201: Local Causal Discovery and Representation Learning: Guided by this causal graph, enhance causal perception on local data, perform a contrastive learning task, and learn local causal feature representations.

[0033] Specifically, This embodiment also provides a causal discovery method based on contrast enhancement: This is a two-way enhancement process. We leverage contrastive learning to improve causal discovery itself. Specifically, by constructing a multi-environment contrastive task, the model compares feature relationships across different data sources, thereby more stably and accurately identifying consistent causal structures across environments. This approach is particularly suitable for addressing the interference of heterogeneous data (such as variations in noise distribution) on traditional causal discovery methods.

[0034] S202: Federated Aggregation and Invariant Feature Extraction: Each client encrypts and uploads the updated model parameters (especially those representing causal relationships) to the server.

[0035] Preferred, This embodiment also provides a causal invariant feature aggregation for federated learning: To achieve bias reduction and generalization in a distributed environment, this method designs a novel federated aggregation strategy. During local training on the client side of the federated learning (e.g., hospitals, factories), each client learns local causal features using the aforementioned technique. On the server side, during aggregation, instead of simply averaging all parameters, it prioritizes causally consistent feature representations shared by all clients and employs residual adaptive parameter aggregation and other schemes to improve communication efficiency and model accuracy.

[0036] S3: Based on the updated model parameters, the server runs a knowledge-guided aggregation algorithm to filter and aggregate features corresponding to stable causal relationships, update the parameters of the global model, and distribute the updated global model to each client. S301: The server runs a knowledge-guided aggregation algorithm to analyze the consistency of causal structures contained in parameters from different clients, filter and aggregate features corresponding to causal relationships that are stable in most sites, and update the global model.

[0037] S302: Global Model Inference and Application: The optimized global model is distributed to various clients or new test sites. For new input data, the model can generate bias-free and generalizable features, and use these features for accurate fault prediction and path identification.

[0038] Preferred, The server uses an aggregation algorithm to analyze the causal consistency inherent in the parameters of each factory. For example, it finds that "bearing vibration leading to shutdown" is consistently present in all factories, while "ambient humidity leading to sensor fluctuations" only occurs in coastal factories. The features generated by the global model will filter out interfering factors such as "ambient humidity," retaining only universal causal features such as "bearing vibration." When a new factory is connected, even if the model has not seen the specific environmental data of that factory, it can still perform accurate fault prediction and path identification based on general causal features.

[0039] S4: Real-time collection of local factory sensor time-series data from different data sources by each client, preprocessing and inputting into the global model to generate feature vectors and perform feature mapping transformation to obtain general features characterizing local equipment faults; the general features are then fed into a pre-built prediction model to generate fault analysis results.

[0040] Specifically, This embodiment also provides an interpretable feature selection mechanism: The final feature selection does not rely on traditional statistical indicators, but rather on a composite scoring mechanism that combines causal contribution and contrast discrimination. The importance of a feature is determined by its contribution to the final causal effect (such as user clicks) and its stability across different contrast views, thus outputting a reliable and interpretable subset of features.

[0041] Specifically, This invention utilizes the concept of invariant learning to guide the model to focus on core features that are stable across environments and domains. By filtering causally consistent feature representations shared by all clients during federated learning aggregation, it eliminates spurious correlation factors specific to each site.

[0042] This embodiment also provides a system for predicting factory equipment failures using any of the above-mentioned methods that integrate causal inference, including: Multiple clients, each acquiring local time-series data from different data sources, inputting it into a trained local model, and generating a preliminary causal graph representing the causal relationships of local device failures; The server is used to initialize the parameters of a pre-built global model for generating generalized features; the local time series data is local factory sensor time series data. The local causal discovery module, located on each client, is used to perform causal perception data augmentation on the local factory sensor time series data based on the initial causal graph. Then, each client performs comparative learning of the local model based on the data augmented local factory sensor time series data to obtain the local causal feature representation, updates the local model parameters based on the causal feature representation, and encrypts and uploads the updated model parameters to the server. The knowledge-guided aggregation module, located on the server side, is used to update the global model by running a knowledge-guided aggregation algorithm based on the updated model parameters, and then distribute the updated global model to each client. The prediction module is used to acquire local factory sensor time-series data collected in real time from different data sources by various clients, input it into the global model, generate universal features that characterize the causal relationship of local equipment failures, and send them into the pre-built prediction model to generate equipment failure analysis results.

[0043] Example 2 This embodiment provides a disease prediction system that integrates causal inference and contrastive learning, including: Multiple clients, each client obtains local disease diagnosis data from different data sources, inputs it into a trained local model, and generates a preliminary causal graph representing the causal relationships of local diseases; The server is used to initialize the parameters of a pre-built global model for generating universal features; the local time series data is local disease diagnosis time series data. The local causal discovery module, located on each client, is used to perform causal perception data augmentation on local disease diagnosis data by each client based on the preliminary causal graph. Then, each client performs comparative learning of the local model based on the data-augmented local disease diagnosis time series data to obtain the local causal feature representation, updates the local model parameters based on the causal feature representation, and encrypts and uploads the updated model parameters to the server. The knowledge-guided aggregation module, located on the server side, is used to update the global model by running a knowledge-guided aggregation algorithm based on the updated model parameters, and then distribute the updated global model to each client. The prediction module is used to acquire local disease diagnosis data from different data sources collected in real time by various clients, input it into the global model, generate universal features that characterize the causal relationship of local diseases, and send it into the pre-built prediction model to generate disease analysis results.

[0044] Specifically, This embodiment also provides a causal-guided method for generating contrastive features: This is the cornerstone of the framework. Instead of randomly generating data-augmented views, we leverage preliminarily identified causal structures (such as parent nodes in a causal graph) to guide data augmentation. For example, in mind mapping analysis, the model identifies key invariant subgraphs related to the disease and augments these structures to generate diverse and causally relevant positive samples. Simultaneously, counterfactual samples can be generated based on the causal model as high-quality negative samples, forcing the model to learn more discriminative features.

[0045] Specifically, The application steps of the system are as follows: Causal diagram guidance: First, identify key brain regions and invariant subgraphs that have a causal relationship with a specific disease (such as Alzheimer's disease).

[0046] Data augmentation implementation: When generating augmented views, the system locks the structure and weights of these key subgraphs, and only perturbs background neural noise or irrelevant brain region connections that do not affect disease diagnosis, ensuring that the features learned by the model are always anchored to the pathological causal chain.

[0047] Counterfactual negative samples: Generate a virtual counterfactual brain map of "if the key brain region had not atrophied", forcing the model to identify the causal discriminative features that truly lead to the disease diagnosis.

[0048] Specifically, the system provided in this embodiment no longer uses traditional random masks or noise additions, but instead uses the parent nodes in the initially identified causal graph as anchor points. By keeping the core causal structure unchanged, intervention is only applied to non-causal interference variables to generate a positive sample view; at the same time, counterfactual samples are generated using the causal model as high-quality negative samples.

[0049] Example 3 This embodiment provides a traffic flow prediction system that integrates causal inference and contrastive learning, including: Multiple clients, each acquiring local time-series data from different data sources, inputting it into a trained local model, and generating a preliminary causal graph representing the causal relationships of local traffic flows; The server is used to initialize the parameters of a pre-built global model for generating generalized features; the local time series data is local traffic flow time series data. The local causal discovery module, located on each client, is used to perform causal perception data augmentation on local traffic flow time series data by each client based on the preliminary causal graph. Then, each client performs comparative learning of the local model based on the data augmented local traffic flow time series data to obtain the local causal feature representation, updates the local model parameters based on the causal feature representation, and encrypts and uploads the updated model parameters to the server. The knowledge-guided aggregation module, located on the server side, is used to update the global model by running a knowledge-guided aggregation algorithm based on the updated model parameters, and then distribute the updated global model to each client. The prediction module is used to acquire local traffic flow time-series data from different data sources collected in real time by various clients, input it into the global model, generate universal features that characterize the causal relationship of local traffic flow, and send it into the pre-built prediction model to generate traffic flow analysis results.

[0050] Specifically, When predicting traffic flow under sudden road conditions, a causal graph is constructed that includes weather, holidays, and road network structure to achieve causal traffic flow prediction. When there is a sudden rainstorm, the model relies on the causal invariant feature of "rainfall -> vehicle speed decrease" rather than just historical statistical correlation to make predictions, ensuring the stability of decision-making under extreme weather conditions.

[0051] Specifically, This embodiment also provides a recommendation system that integrates causal inference and contrastive learning to solve the problem of spurious relevance caused by the confounding variable of "popularity bias".

[0052] First, two paths are identified: "user's real interest -> click" and "item popularity -> click". Then, through causal bias correction, the model can finally identify and control popularity factors, effectively mitigating the "Matthew effect" and allowing long-tail and niche content to receive fair exposure.

[0053] This invention proposes a unified data engineering framework that enhances system performance and interpretability throughout the entire process from data preprocessing to model training through the synergistic effect of three core technologies: causal guidance, contrastive learning, and federated aggregation. The framework's design philosophy is to leverage prior knowledge provided by causal structures to guide data augmentation and feature selection, extract cross-domain invariant representations through contrastive learning, and integrate multi-source data through a federated learning mechanism to achieve bias removal and generalization in high-dimensional heterogeneous environments. The proposed method not only significantly improves the accuracy and robustness of decision-making but also addresses the needs for data security and deployment flexibility.

[0054] Therefore, this patented technology aims to construct a unified data engineering and mining framework to systematically solve the following problems: 1) Fundamental problem: Causal confounding. By introducing causal structures into feature generation and selection, potential confounding factors on the user and item sides can be identified and controlled, thus achieving debiasing at the data representation level.

[0055] 2) Key obstacle: Out-of-distribution generalization. By utilizing the concept of invariant learning in causal inference, we guide contrastive learning to focus on core features that are stable across environments and domains, thereby improving the robustness of the model in unknown scenarios.

[0056] 3) Practical Constraints: Data Privacy and Heterogeneity. Combining the federated learning architecture, a distributed training mechanism suitable for causal and contrastive learning is designed to generate and filter global features while ensuring that data does not leave the local machine.

[0057] The beneficial effects of this invention are: The following beneficial effects are expected to be achieved by implementing this technical solution: - More reliable decision-making: In the medical field, it can help doctors identify the true key biomarkers of diseases; in the industrial field, it can clearly show the propagation path of faults, helping to facilitate precise operation and maintenance.

[0058] - A fairer system: The recommendation system can effectively mitigate the "Matthew effect," giving long-tail products and niche content a fairer exposure opportunity and improving the health of the ecosystem.

[0059] - More flexible deployment: Thanks to its natural integration with federated learning, the framework can build joint models across multiple institutions without aggregating sensitive data, accelerating research processes in areas such as drug efficacy evaluation.

[0060] - More robust performance: When the data distribution changes (such as when a recommendation system encounters new users or traffic prediction encounters sudden road conditions), the model can exhibit stronger adaptability and stability by relying on causal invariant features.

[0061] Application prospects of this invention: This framework has the potential to enable a new generation of intelligent solutions in several key areas: Smart healthcare: Building a federated network for disease diagnosis across hospitals to jointly discover more universal etiological patterns and biomarkers while protecting patient privacy.

[0062] Industrial Internet: Enables predictive maintenance across factories and equipment, accurately identifies fault propagation paths, and greatly improves production safety and efficiency.

[0063] The next generation of recommendation systems: creating a smarter, fairer recommendation engine that can break the "information cocoon" and better meet users' true interests through causal bias correction.

[0064] Intelligent Transportation: While protecting the privacy of traffic data in various regions, we will build a more macroscopic and accurate traffic flow prediction model to provide decision support for the city's brain.

Claims

1. A factory equipment failure prediction method integrating causal inference, characterized in that, include: Step 1: Each client obtains local time-series data from different data sources, inputs it into the trained local model, and generates a preliminary causal graph representing local device faults; A pre-built global model for extracting general features from input data is initialized by the server; the local time-series data is local factory sensor time-series data. Step 2: Based on the preliminary causal graph, each client performs causal perception data augmentation on the local factory sensor time series data; then, each client performs comparative learning of the local model based on the data-augmented local factory sensor time series data to obtain the local causal feature representation, updates the local model parameters based on the causal feature representation, and encrypts and uploads the updated model parameters to the server. Step 3: Based on the updated model parameters, the server runs a knowledge-guided aggregation algorithm to filter and aggregate features corresponding to stable causal relationships, update the parameters of the global model, and distribute the updated global model to each client. Step 4: Each client collects local factory sensor time-series data from different data sources in real time, preprocesses it, and inputs it into the global model to generate feature vectors and perform feature mapping transformation to obtain general features characterizing local equipment faults; the general features are then sent into a pre-built prediction model to generate fault analysis results.

2. The factory equipment failure prediction method integrating causal inference according to claim 1, characterized in that, Each client uses the BFS causal discovery method to train the local model, then inputs the local factory sensor time series data, and the trained local model outputs a preliminary causal graph representing the faults of the local equipment.

3. The factory equipment failure prediction method integrating causal inference according to claim 1, characterized in that, The server runs a knowledge-guided aggregation algorithm to analyze the consistency of causal structures contained in parameters from different clients, filter and aggregate features corresponding to stable causal relationships in the factory clients, and update the global model.

4. The factory equipment failure prediction method integrating causal inference according to claim 3, characterized in that, The residual adaptive parameter aggregation method is used to aggregate the features corresponding to the causal associations; and when aggregating the features corresponding to the causal associations of each client, the feature representations that are shared by each client and are causally consistent are selected first.

5. The factory equipment failure prediction method integrating causal inference according to claim 1, characterized in that, Step three also includes: After the updated global model is distributed to each client, new local factory sensor time-series data is input into the global model. The global model generates data features that are universally representative of local equipment faults, and subsequent fault prediction and path identification are performed based on the data features.

6. The factory equipment failure prediction method integrating causal inference according to claim 5, characterized in that, The importance of the data features is evaluated based on a composite scoring mechanism that combines causal contribution and contrast discrimination, and the final data features with universality are selected. The importance of the data features is mainly determined by their contribution to the final causal effect and their stability under different contrast views.

7. The factory equipment failure prediction method integrating causal inference according to claim 1, characterized in that, The parameter update process for each client in step three includes: After the updated global model is distributed to each client, fault prediction and path identification data corresponding to the newly collected local factory sensor time series data are collected, and it is determined whether the local model parameters need to be updated. If the pre-set fault prediction and path identification indicators are met, training is not required. If the fault prediction and path identification indicators are not met, the process returns to step two.

8. The factory equipment failure prediction method integrating causal inference according to claim 1, characterized in that, The data enhancement for causal perception specifically includes: The data augmentation is guided by the parent nodes in the preliminary causal graph; the local model identifies and generates a key invariant subgraph, and then performs data augmentation based on the key invariant subgraph to generate diverse and causally related positive samples and counterfactual negative samples, which are then used for comparative learning of the local model.

9. A factory equipment failure prediction method integrating causal inference according to claim 8, characterized in that, The negative samples are virtual counterfactual sample data without causal correlation generated based on the key invariant subgraph.

10. A system for predicting factory equipment failures using a method incorporating causal inference as described in any one of claims 1-9, characterized in that, include: Multiple clients, each acquiring local time-series data from different data sources, inputting it into a trained local model, and generating a preliminary causal graph representing the causal relationships of local device failures; The server is used to initialize the parameters of a pre-built global model for generating generalized features; the local time-series data is local factory sensor time-series data. The local causal discovery module, located on each client, is used to perform causal perception data augmentation on the local factory sensor time series data based on the preliminary causal graph. Then, each client performs comparative learning of the local model based on the data-augmented local factory sensor time series data to obtain the local causal feature representation, updates the local model parameters based on the causal feature representation, and encrypts and uploads the updated model parameters to the server. The knowledge-guided aggregation module, located on the server side, is used to update the global model by running a knowledge-guided aggregation algorithm based on the updated model parameters, and then distribute the updated global model to each client. The prediction module is used to acquire local factory sensor time-series data collected in real time from different data sources by various clients, input the data into the global model, generate universal features characterizing the causal relationship of local equipment failures, and send them into the pre-built prediction model to generate equipment failure analysis results.

Images