Equipment fault prediction method based on industrial causal logic

Abstract

The invention discloses an equipment fault prediction method based on industrial causal logic. The method comprises the following steps: acquiring an unbalanced equipment monitoring data set; missing value processing is carried out on the unbalanced equipment monitoring data set, and a causal relation graph is established; quantifying a causal relationship and calculating a causal weight matrix; constructing a causal constraint generative adversarial network, taking a causal weight matrix as an attention weight to generate a synthetic fault sample, combining the synthetic fault sample with an original fault sample to form a balanced fault sample set, and finally combining the balanced fault sample set with a normal sample set to form a balanced data set; and training a classifier based on the balanced data set and outputting a fault prediction result. According to the method, by identifying and utilizing the physical causal relationship in the equipment data, it is ensured that the generated synthetic equipment state sample strictly follows the engineering logic, unreasonable engineering samples generated by a traditional method are avoided, the accuracy of equipment fault prediction is remarkably improved, and therefore the equipment shutdown loss caused by false detection and missing detection is reduced.

Description

Technical Field

[0001] This invention belongs to the technical field of industrial equipment failure prediction methods, specifically relating to an equipment failure prediction method based on industrial causal logic. Background Technology

[0002] Equipment failure is a type of industrial event with an extremely low probability of occurrence but serious consequences, such as bearing failure in mechanical equipment, short circuits in motor windings, and leaks in hydraulic systems. Because equipment failure events are few in number, typically accounting for only 1% to 5% of total operating data, the sample data used to train fault diagnosis models is severely imbalanced. If equipment failures are not detected early, many pieces of equipment will miss the optimal maintenance window due to delayed repairs, leading to high downtime and equipment damage.

[0003] Existing equipment failure prediction methods mainly employ traditional machine learning algorithms such as support vector machines, decision trees, and random forests. However, existing industrial datasets suffer from the following technical problems: Problem 1: Severe data imbalance. For example, in a bearing failure dataset collected by a factory, the failure samples account for only 1%, resulting in an imbalance ratio of 100:1, making it difficult for the model to accurately identify minority class samples.

[0004] Problem 2: Inefficient sample generation. Traditional oversampling methods such as SMOTE generate new samples by linear interpolation between existing minority class samples, but this can produce equipment state samples that are not engineered to fit the requirements, such as equipment with very low rotation speed but high temperature.

[0005] Problem 3: Ignoring industrial causal relationships. Industrial data contains abundant physical causal relationships, such as rotational speed → vibration → temperature → failure risk, but existing generation methods lack causal constraint mechanisms and cannot ensure that the generated samples conform to engineering logic.

[0006] For example, W. Wu et al. used the SVM-RFE method to process industrial equipment datasets. This method reduces computational cost through recursive feature elimination, but it still has significant problems when dealing with extremely imbalanced data. First, this method mainly relies on statistical correlation in the feature selection process, ignoring the physical causal relationships that are widespread in the industrial field. Second, when dealing with minority class samples, the statistical features are often unstable due to the small sample size, leading to reduced reliability of feature selection. Finally, this method cannot solve the fundamental problem of data imbalance; it can only perform feature selection based on existing imbalanced data.

[0007] While the ensemble method developed by R. Alsmariy et al. performs well on some evaluation metrics, achieving an ROC-AUC value of 98.49%, in-depth analysis reveals serious engineering credibility issues. This method employs the SMOTE algorithm for data augmentation, generating new samples through linear interpolation between minority class samples. However, this simplistic mathematical interpolation method completely ignores common-sense engineering principles and physical causal relationships. For example, interpolating a healthy device operating under low load with a faulty device operating under high load might generate a virtual device state with moderate load and moderate failure risk. However, the combination of operating parameters in this device state might be completely unreasonable from an engineering perspective, such as an abnormal state of low speed combined with high temperature.

[0008] Furthermore, existing methods have the following problems when dealing with multidimensional industrial features: The assumption of feature independence is flawed: Traditional methods typically assume that the parameters of various devices are independent of each other, but in reality, complex interdependencies exist in industrial data. For example, there are clear physical causal chains between indicators such as equipment speed, vibration, temperature, and power consumption. Ignoring these relationships will lead to the generated samples being unreliable in engineering applications.

[0009] Ignoring temporal causal relationships: Equipment failures often exhibit a clear temporal sequence. For example, bearing failure typically progresses through a process of slight vibration → increased vibration → temperature rise → severe failure. Existing methods cannot capture such temporal causal relationships, and the generated samples may exhibit reversed causal relationships.

[0010] Lack of multi-level causal structure: Industrial data contains both direct causal relationships (e.g., load directly affects power consumption) and indirect causal relationships (e.g., load indirectly affects bearing life by influencing temperature). Existing methods cannot distinguish and process this multi-level causal structure.

[0011] Insufficient integration of domain knowledge: Existing methods mainly rely on data-driven statistical learning, lacking effective integration of engineering domain knowledge. The maintenance experience of equipment experts and the causal relationship findings in engineering literature cannot be effectively incorporated into the data augmentation process.

[0012] To address the aforementioned technical issues, researchers both domestically and internationally have conducted some exploratory work. Zhang et al. proposed a data augmentation method based on device ontology, attempting to use device knowledge graphs to guide sample generation. However, this method mainly focuses on the feature selection stage and fails to fundamentally solve the problem of causal consistency in generated samples. Liu et al. developed constrained generative adversarial networks for device monitoring data augmentation. Although it achieved good results in the field of time-series data, its constraint mechanism mainly targets the continuity of time series and cannot be directly applied to causal constraints for tabular device data.

[0013] In-depth analysis reveals that the core technical problem is that existing data augmentation methods cannot ensure that the generated equipment status samples conform to engineering causal logic while addressing the imbalance in industrial data, resulting in low prediction accuracy and poor engineering reliability. Specifically, this manifests as follows: Statistically reasonable but engineering-inefficient: The generated sample is statistically similar to the real data, but violates basic engineering common sense and physical laws.

[0014] Broken causal relationship: Ignoring the causal dependencies between features leads to inconsistent causal logic in the feature combinations of generated samples.

[0015] Engineering uninterpretability: The lack of engineering interpretability means that equipment experts cannot understand or trust the model's predictions.

[0016] Poor generalization ability: It performs poorly on data from new industrial scenarios or different factories and lacks robustness. Summary of the Invention

[0017] The purpose of this invention is to provide a method for predicting equipment failures based on industrial causal logic, which solves the problem that existing methods generate equipment state samples that do not conform to engineering causal logic, resulting in low prediction accuracy and poor engineering reliability.

[0018] The technical solution adopted in this invention is: a device failure prediction method based on industrial causal logic, comprising the following steps: Step 1: Obtain the monitoring dataset of unbalanced equipment; Step 2: Handle missing values ​​in the unbalanced equipment monitoring dataset and establish a causal relationship graph; Step 3: Quantify causal relationships and calculate the causal weight matrix; Step 4: Construct a causal constraint generative adversarial network, use the causal weight matrix as attention weights to generate synthetic fault samples, then merge them with the original fault samples to form a balanced fault sample set, and finally merge the balanced fault sample set with the normal sample set to form a balanced dataset. Step 5: Train a classifier based on the balanced dataset and output the fault prediction results.

[0019] The invention is further characterized in that, Step 1 includes the following steps: Step 1.1: Synchronously collect basic equipment information, real-time operating parameters, multi-dimensional sensor data, historical maintenance records, and fault markers through the industrial data acquisition system; Step 1.2: Unify the fault markers by encoding "fault marker = 0" to indicate normal equipment and "fault marker = 1" to indicate faulty equipment; Step 1.3: Align all collected data according to timestamps to form an unbalanced equipment monitoring dataset as the original fault sample.

[0020] Step 2 includes the following steps: Step 2.1: Identify missing values ​​in the unbalanced equipment monitoring dataset, delete sensor features with a missing ratio ≥ 50%, and fill in features with a missing ratio < 50% using engineering rationality. Step 2.2: Input the dataset after missing values ​​processing into the causal discovery algorithm to identify the causal dependencies between device operating parameters, and output the identification results in the form of a directed acyclic graph to obtain the causal relationship graph.

[0021] The imputation method in step 2.1 is as follows: for continuous operating parameters with a missing ratio of <50%, K-nearest neighbor imputation based on similar equipment operating conditions is used; for discrete state features with a missing ratio of <50%, rule reasoning based on decision trees is used; for time-series sensor data with a missing ratio of <50%, sequence prediction imputation is used using long short-term memory networks; for parameter combinations with physical constraints, constraint optimization imputation is used to ensure that the imputed values ​​conform to engineering reality.

[0022] Step 3 includes the following steps: Step 3.1: Perform partial correlation coefficient analysis, mutual information calculation, Granger causality test and transfer entropy analysis on each pair of variables in the causal relationship diagram in sequence to obtain the strength value of each causal relationship and filter out weak correlation relationships; Step 3.2: Standardize, encode one-hot, and normalize the numerical, categorical, and time-series features respectively; Step 3.3: Use the following fusion formula to fuse the results of multiple methods with the domain expert knowledge weights, and output a causal weight matrix. :

[0023] in, Indicates the first i The weight matrix obtained by the causal analysis method Indicates the corresponding fusion weights. Represents the knowledge weight of domain experts. β The weighting coefficients representing expert knowledge. k This indicates the number of causal analysis methods.

[0024] Step 3.1 specifically involves: First, partial correlation coefficients are used to analyze the linear causal strength between operating parameters and fault development. The calculation formula is as follows:

[0025] in, Represents control variables Z back X andY The partial correlation coefficient, express X and Y The simple correlation coefficient, Z Indicates control variables; Then, the mutual information method is used to evaluate the strength of nonlinear causality. :

[0026] in, Representing variables X and Y The mutual information between variables is used to assess the strength of the nonlinear causal relationship between the two variables. express X and Y The joint probability distribution of ; Representing variables X The marginal probability distribution; Representing variables Y The marginal probability distribution; Next, the Granger causality test was used to assess the temporal causality. :

[0027] in, Indicates from variable X to variable Y The Granger causality test statistic follows a certain order. F Distribution, used for testing X Is it correct? Y It exhibits Granger causality; To represent the sum of squared residuals of the restricted model, use only... Y Historical value prediction Y The future value; Represents the sum of squared residuals of an unrestricted model, and uses X and Y Historical value prediction Y The future value; p Indicates the lag order. n Indicates the total number of samples; Next, the transitive entropy method is used to evaluate the strength of temporal causality. :

[0028] in, express At any moment of 3D historical state vector express At any moment of 3D historical state vector and These are the embedding dimensions, Represent the corresponding probability density function, through Nearest neighbor method estimation; Finally, relationships with low causal strength, i.e., threshold |r|<0.2, were selected and filtered.

[0029] Step 4 includes the following steps: Step 4.1: Construct a causal constraint generator. The input layer receives 100-dimensional random noise, which sequentially passes through a 256-dimensional fully connected layer, a causal attention layer, a 128-dimensional fully connected layer, and a 32-dimensional output layer to output a feature vector. The causal attention layer incorporates the causal weight matrix obtained in Step 3 using the following formula. Constrained output feature vector:

[0030] in, Q , K , V These represent the query, key, and value matrices, respectively. α This represents the causal constraint strength parameter; simultaneously, position encoding is added using the following formula:

[0031]

[0032] in, pos For location, i For dimensional indexing, For the embedding dimension of the model; Step 4.2: Construct a multi-task discriminator. The input layer receives a 32-dimensional feature vector, which is then passed through a 128-dimensional fully connected layer, a Dropout layer, a 64-dimensional fully connected layer, and a 2-dimensional output layer to output the results. The discriminator calculates the physical constraint violation loss using the following formula. :

[0033] in, Indicates the first i A physical constraint function, To constrain tolerance, m To constrain the total number, Tolerance / threshold; and adversarial training with the generator, using the WGAN-GP loss function; Step 4.3: Calculate the adversarial losses separately. Loss of causal consistency Diversity loss and feature matching loss Among them, combating losses The calculation formula is:

[0034] in, This indicates that the discriminator is effective against real samples. The judgment result, The generator is based on noise. The generated samples, where 𝔼 represents the mathematical expectation, Represents the true data distribution. Indicates noise distribution; Loss of causal consistency The calculation formula is:

[0035] in, This represents the relevance matrix of the generated samples. A causal weight matrix representing real data. d Indicates feature dimension; Diversity loss Calculated based on sample similarity within a batch:

[0036] in, B For batch size, , All are generated samples. ε For numerical stability parameters; Feature matching loss The calculation formula is:

[0037] in, Indicates the first The weighting coefficients of each project constraint. Indicates the first Each engineering constraint function This represents the state sample of the synthesis equipment generated by the generator. This represents a sample of real equipment operating data. This indicates the total number of engineering constraints. This represents the square of the L2 norm, used to calculate the degree of difference; The total loss is then obtained by weighted summation of the four types of losses using the following formula. Backpropagation updates the generator and discriminator parameters:

[0038] Among them, the weighting coefficient , , , The constraint strength parameters are determined using a grid search method based on the validation set performance. α Adaptive update via the following formula:

[0039] in, η For learning rate, For the updated strength parameters, Current constraint strength parameters To The gradient; Step 4.4: After training is completed, generate a number of synthetic fault samples equal to the number of normal samples. Merge the synthetic fault samples with the original fault samples to form a balanced fault sample set. Then merge the balanced fault sample set with the normal sample set to form a balanced dataset. Perform stratified sampling on the balanced dataset to obtain the training set and the test set.

[0040] Step 5 includes the following steps: Step 5.1: Input the training set into the Support Vector Machine (SVM) for training. The SVM fault diagnosis model is obtained through training using the following formula:

[0041] in, X Represents equipment characteristic data, W Represents the weight vector. b Indicates the bias term. f ( x Used to determine if the equipment is faulty; Step 5.2: Input the training set into the Random Forest for training, and obtain the RF fault diagnosis model through the following formula:

[0042] in, I It is an indicator function. The prediction result of the m-th classifier For real labels, if ,but I =1; otherwise I =0, M Indicates the number of decision trees; Step 5.3: Input the test set into the SVM fault diagnosis model and the RF fault diagnosis model respectively, and output the fault prediction results for each. Step 5.4: Use a voting mechanism to merge the fault prediction results of the SVM fault diagnosis model and the RF fault diagnosis model to obtain the final fault prediction result.

[0043] The beneficial effects of this invention are: the equipment failure prediction method based on industrial causal logic of this invention, by identifying and utilizing the physical causal relationships in equipment data, ensures that the generated synthetic equipment state samples strictly follow engineering logic, avoids unreasonable engineering samples generated by traditional methods, significantly improves the accuracy of equipment failure prediction, and thus reduces equipment downtime losses caused by false detections and missed detections. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the equipment failure prediction method based on industrial causal logic of the present invention. Figure 2 This is a detailed architecture diagram of the causal discovery system in the equipment failure prediction method based on industrial causal logic of the present invention; Figure 3 This is a flowchart of the advanced causal weight calculator in the equipment failure prediction method based on industrial causal logic of this invention; Figure 4 This is a network structure diagram of the advanced causal attention mechanism in the equipment failure prediction method based on industrial causal logic of this invention; Figure 5 This is an improved generative adversarial network architecture diagram in the equipment failure prediction method based on industrial causal logic of this invention; Figure 6 This is a flowchart illustrating the composition and calculation of the multi-objective loss function in the equipment failure prediction method based on industrial causal logic of this invention. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0046] This invention provides a method for equipment failure prediction based on industrial causal logic, specifically addressing the technical problems of severely imbalanced data, unreasonable sample generation engineering, and neglect of physical causal relationships in equipment failure prediction. The method first acquires an imbalanced equipment monitoring dataset and performs missing value processing and causal relationship discovery. Then, it quantifies and weights the causal relationships. Next, it uses a generative adversarial network incorporating causal constraints to generate equipment state samples that conform to engineering logic. Finally, it achieves accurate prediction through classifier training. Specifically, it is reflected in: 1) This invention establishes a three-dimensional causal analysis framework encompassing direct causal relationships, indirect causal relationships, and temporal causal relationships through hierarchical causal relationship mining technology. Direct causal relationships include "load rate → power consumption," indirect causal relationships include "load rate → temperature → lubrication effect → failure risk," and temporal causal relationships include "current vibration → future temperature → subsequent failure probability." Through a causal discovery engine that integrates multiple algorithms, it can automatically identify up to 50 different types of causal relationship patterns in complex industrial systems.

[0047] 2) This invention employs an adaptive causal constraint mechanism to dynamically adjust the strength of causal constraints, automatically adjusting constraint parameters based on different equipment types, operating conditions, and fault types. This mechanism continuously optimizes the constraint strategy through reinforcement learning algorithms, ensuring that generated samples maintain diversity while strictly adhering to physical laws. (Constraint strength parameters) α The following adaptive update formula is adopted:

[0048] in, η For learning rate, For the comprehensive loss function, For the updated strength parameters, Current constraint strength parameters To The gradient.

[0049] 3) This invention constructs a collaborative optimization framework integrating four objectives: generation quality, causal consistency, sample diversity, and engineering rationality, through a multi-objective collaborative optimization generator. The generator must not only pass the discriminator's true / false judgment but also satisfy causal consistency constraints, maintain sufficient differences between samples, and be verified by the engineering rationality evaluation module. The total loss function is designed as follows:

[0050] The weight coefficients are determined using a grid search method.

[0051] 4) This invention employs an engineering knowledge-enhanced discriminator design, incorporating expert knowledge to create a multi-task discriminator that not only judges the statistical authenticity of samples but also assesses engineering rationality and physical consistency. The discriminator adopts a hierarchical architecture: the bottom layer extracts basic statistical features, the middle layer identifies violations of physical laws, and the top layer performs a comprehensive evaluation. Through real-time interaction with the expert knowledge base, the discriminator continuously improves its accuracy.

[0052] 5) This invention proposes a causal-aware attention mechanism specifically for industrial time-series data. This mechanism can automatically adjust the attention weight allocation based on the causal relationship graph, prioritizing feature combinations with strong causal relationships. The attention weight calculation incorporates temporal information and physical constraints.

[0053] in This is the time-series weight matrix. β This refers to the temporal constraint strength parameter.

[0054] This invention provides a method for predicting equipment failures based on industrial causal logic, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the unbalanced equipment monitoring dataset. This dataset contains multi-level feature information such as basic equipment information, real-time operating parameters, multi-dimensional sensor data, historical maintenance records, and fault markers. It is obtained synchronously from multiple monitoring points through an industrial data acquisition system to ensure data integrity and time consistency. Fault markers use '0' to represent normal equipment and '1' to represent faulty equipment. The proportion of faulty equipment is usually less than 5%.

[0055] Step 2: Perform missing value processing and causal relationship discovery on the acquired device dataset to obtain a causal relationship graph. This step automatically detects data quality issues using an intelligent missing value identification algorithm, employs an imputation strategy based on the device's physical model, and utilizes a combination of multiple causal discovery algorithms to identify complex causal dependencies between device operating parameters. Specifically, it includes the following steps: Step 2.1: Process missing values ​​in the acquired device dataset. First, identify the missing values ​​in the dataset. Second, delete sensor features with more than 50% missing data. Finally, fill in the missing data for features with less than 50% missing data using engineering principles.

[0056] Data imputation methods include: for operational parameter features, imputation is performed using conditional mean based on basic information such as equipment model and operating conditions; for state features, intelligent imputation is performed based on equipment operating patterns. Specific imputation strategies are as follows: for continuous operational parameters, the K-nearest neighbor imputation method based on similar equipment operating conditions is used; for discrete state features, rule-based reasoning based on decision trees is used; for time-series correlated sensor data, sequence prediction imputation is performed using Long Short-Term Memory (LSTM) networks; for parameter combinations with clear physical constraints, constraint optimization methods are used to ensure that the imputed values ​​conform to engineering realities.

[0057] Step 2.2: Use causal discovery algorithms to identify physical causal relationships in device data and construct a causal relationship graph for fault development. The causal relationship graph represents the causal dependencies between various operating parameters in the form of a directed acyclic graph. The detailed architecture diagram of the causal discovery system is shown below. Figure 2 As shown.

[0058] Step 3: Quantitatively analyze and calculate the weights of causal relationships in the equipment data to obtain a causal weight matrix. This process employs a multi-dimensional causal strength assessment method, combining partial correlation analysis, mutual information calculation, Granger causality tests, and other statistical methods to comprehensively evaluate the causal strength between each pair of variables and construct a high-precision causal weight matrix. The workflow diagram of the advanced causal weight calculator is shown below. Figure 3 As shown. Specifically, it includes the following steps: Step 3.1: Calculate the strength of causal relationships in the equipment data using a multi-dimensional causal strength assessment framework, including partial correlation coefficient analysis, mutual information calculation, Granger causality test, and transit entropy analysis. First, use partial correlation coefficient analysis to determine the linear causal strength between operating parameters and fault development. The calculation formula is:

[0059] in, Represents control variables Z back X and Y The partial correlation coefficient, express X and Y The simple correlation coefficient, Z Indicates a control variable.

[0060] Then, the mutual information method is used to evaluate the strength of nonlinear causality:

[0061] in, Representing variables X and Y The mutual information between variables is used to assess the strength of the nonlinear causal relationship between the two variables. express X and Y The joint probability distribution of ; Representing variables X The marginal probability distribution; Representing variables Y The marginal probability distribution.

[0062] Next, the Granger causality test is used to assess the temporal causality:

[0063] in, Indicates from variable X to variable Y The Granger causality test statistic follows a certain order. F Distribution, used for testing X Is it correct? Y It exhibits Granger causality; To represent the sum of squared residuals of the restricted model, use only... Y Historical value prediction Y The future value; Represents the sum of squared residuals of an unrestricted model, and uses X and Y Historical value prediction Y The future value; p Indicates the lag order.n This represents the total number of samples.

[0064] Next, the transitive entropy method is used to evaluate the strength of temporal causality. :

[0065] in, ,from X arrive Y Transmission entropy, a measure X right Y The strength of causal influence express At any moment of 3D historical state vector express At any moment of 3D historical state vector and These are the embedding dimensions, Represent the corresponding probability density function, through Nearest neighbor method estimation.

[0066] Finally, relationships with low causal strength (i.e., |r|<0.2) were selected and filtered to construct a simplified causal graph containing strong causal relationships.

[0067] Step 3.2: Perform multi-level standardization on the features in the device data, including Z-score standardization for numerical features, one-hot encoding for categorical features, and sliding window normalization for time-series features. The Z-score standardization formula is as follows:

[0068] Where X represents device characteristic data. σ represents the mean of the feature data, and σ represents the standard deviation of the feature data. This represents the standardized data. A feature weight vector is also established:

[0069] Each weight The importance of this feature in the causal graph is determined.

[0070] Step 3.3: Construct a multi-dimensional causal weight fusion matrix, which fuses the weights obtained from different causal analysis methods. The fusion formula is as follows:

[0071] in, Indicates the first iThe weight matrix obtained by the causal analysis method Indicates the corresponding fusion weights. Represents the knowledge weight of domain experts. β The weighting coefficients representing expert knowledge. k This indicates the number of causal analysis methods.

[0072] Step 4: Use a generative adversarial network based on causal constraints to generate samples from the device data, such as... Figure 5 As shown. This network architecture innovatively integrates, for example, Figure 4 The causal attention mechanism, multi-objective loss function, and adaptive constraint regulator shown ensure that the generated samples meet both statistical distribution requirements and engineering physical constraints through an adversarial training process. The generated data is then integrated and split into training and test sets using stratified sampling. Specifically, the following steps are included: Step 4.1: Construct a causal constraint generator. This generator employs a multi-layer neural network architecture, incorporating engineering causal constraints to ensure the generation process follows discovered physical causal relationships. The generator includes a noisy input layer, a causal embedding layer, a causal attention layer, and a feature output layer. The causal attention mechanism formula is:

[0073] in, Q , K , V These represent the query, key, and value matrices, respectively. Represents the causal weight matrix. α This represents the parameter indicating the strength of the causal constraint. A positional encoding mechanism is also introduced.

[0074]

[0075] in pos For location, i For dimensional indexing, This represents the embedding dimension of the model.

[0076] Step 4.2: Construct a multi-task discriminator. This discriminator not only evaluates the authenticity of the generated samples but also performs multi-dimensional evaluations of engineering rationality and physical consistency. The discriminator adopts a hierarchical architecture, including a feature extraction layer, a physical constraint verification layer, and a comprehensive evaluation layer. The physical constraint verification is performed by establishing a constraint violation loss function:

[0077] in Indicates the first i A physical constraint function, To constrain tolerance, mTo constrain the total number, The tolerance / threshold is used. The quality of generated samples is optimized through adversarial training, and the WGAN-GP loss function is employed to ensure training stability.

[0078] Step 4.3: Design a multi-objective loss function, such as... Figure 6 As shown. Ensuring the generated device state samples follow engineering causal logic includes four components: adversarial loss, causal consistency loss, diversity loss, and feature matching loss.

[0079] Combat losses The calculation formula is:

[0080] in, This indicates that the discriminator is effective against real samples. The judgment result, The generator is based on noise. The generated samples, where 𝔼 represents the mathematical expectation, Represents the true data distribution. The loss function represents the noise distribution and is used to train the adversarial process between the generator and the discriminator.

[0081] The formula for calculating causal consistency loss is:

[0082] in, This represents the relevance matrix of the generated samples. A causal weight matrix representing real data. d Indicates the feature dimension.

[0083] Diversity loss is calculated using in-batch sample similarity:

[0084] in B For batch size, , All are generated samples. ε This is a numerical stability parameter.

[0085] Feature matching loss The calculation formula is:

[0086] in, This represents the loss function for engineering rationality. Indicates the first The weighting coefficients of each project constraint. Indicates the first Each engineering constraint function is used to test specific physical laws. This represents the state sample of the synthesis equipment generated by the generator. This represents a sample of real equipment operating data. This indicates the total number of engineering constraints. This represents the square of the L2 norm, used to calculate the degree of difference.

[0087] Step 4.4: Using a progressive training strategy and adaptive learning rate scheduling, the generated synthetic samples and the original data are merged into a complete balanced dataset after quality assessment and intelligent screening. Then, a hierarchical cross-validation method is used to split the dataset into training and test sets to ensure that the ratio of faulty samples to normal samples in each subset remains consistent, while ensuring that the continuity of time series data is not disrupted.

[0088] Step 5: Input the training set into multiple classifiers for ensemble training, including support vector machines, random forests, and deep neural networks. The prediction results of multiple classifiers are fused through a voting mechanism, and then predictions are made on the test set to obtain the final fault prediction result and confidence assessment. The classifier training specifically involves: The training set is input into a Support Vector Machine (SVM) for classification training, and then predictions are made on the test set. The SVM decision function is expressed as:

[0089] in, X Represents equipment characteristic data, W Represents the weight vector. b Indicates the bias term. f ( x This is used to determine if the equipment is faulty.

[0090] The training set is input into a Random Forest (RF) for classification training, and then predictions are made on the test set. The RF prediction formula is:

[0091] in, I It is an indicator function. The prediction result of the m-th classifier For real labels, if ,but I =1; otherwise I =0, M Indicates the number of decision trees.

[0092] Through the above methods, the equipment failure prediction method based on industrial causal logic provided by this invention systematically solves the data imbalance problem in equipment failure prediction through five core technical steps, and ensures the engineering rationality and physical consistency of the generated samples. This invention first acquires an imbalanced equipment monitoring dataset, which includes basic equipment information, real-time operating parameters, multi-dimensional sensor data, and multi-level feature information such as fault markers. The fault markers use binary encoding, with '0' representing normal equipment status and '1' representing faulty equipment status. Due to the rarity and randomness of industrial equipment failure events, the proportion of faulty equipment in the entire dataset is usually less than 5%. This extreme class imbalance severely affects the predictive performance of traditional machine learning algorithms, causing the model to tend to predict all samples as the majority class, resulting in high false positive and false negative rates.

[0093] Next, the acquired equipment dataset undergoes intelligent missing value processing and deep causal relationship discovery to obtain a complete causal relationship graph. This step employs a deep causal discovery framework based on variational inference, using an innovative graph neural network structure learning algorithm to automatically discover complex nonlinear causal dependencies between equipment operating parameters. This Bayesian structure learning method can automatically identify multi-level physical causal chains such as "load rate → power consumption → temperature rise → bearing wear → failure risk" in massive industrial data. At the same time, the framework also integrates a deep learning-based intelligent missing value imputation mechanism. For sensor features with more than 50% missing data, an automatic deletion strategy is adopted, while for key features with less than 50% missing data, a reasonable imputation method based on equipment physical models and engineering experience is used to ensure that the imputed data satisfies both statistical consistency and conforms to actual engineering conditions.

[0094] Then, the causal relationships identified in the equipment data are precisely quantified and dynamically weighted to obtain a high-precision causal weight matrix. This process uses an attention-based meta-learning framework for adaptive weight estimation, and the meta-learning algorithm enables rapid adaptation and accurate estimation of the strength of causal relationships in different industrial scenarios.

[0095]

[0096] The parameters of the model at step t. The loss gradient on the support set; MetaLearner: a meta-learner (such as LSTM or neural networks) that takes the current parameters and gradient and outputs a weight vector. (The learning rate used to adjust the gradient of each component of the query set).

[0097] Simultaneously, a unified model of complex causal relationships across multiple dimensions and time series is constructed by combining a causal strength assessment model based on higher-order tensor decomposition.

[0098] in Higher-order causal tensors can capture the nonlinear interactions between device parameters. The eigenvalues ​​reflect the importance of each causal pattern. The feature vector of the nth pattern is used to represent causal features in different dimensions. R The tensor rank determines the complexity and expressive power of the model. This high-dimensional tensor decomposition method can effectively handle complex causal relationships in industrial data, such as multivariate coupling, temporal dependence, and nonlinear interaction, providing precise causal constraint guidance for the subsequent generation process.

[0099] The core technology involves using a generative adversarial network based on deep variational causal constraints to generate intelligent samples from device data. This innovative network architecture integrates a variational autoencoder, a graph convolutional neural network, and an adversarial training mechanism. First, the variational encoder maps the input device state data to a low-dimensional latent space.

[0100] in, It is an approximate posterior distribution, usually assumed to be a Gaussian distribution. and : These are the mean and variance of the encoder network output.

[0101] Furthermore, a variational lower bound loss function is employed to ensure the continuity and interpretability of the latent representation:

[0102] The first term is the reconstruction loss (negative log-likelihood), and the second term is the KL divergence (regularization term). It is the weight parameter (β-VAE) that controls disentanglement.

[0103] Next, graph convolutional networks are used to perform information propagation and feature learning on the causal graph structure:

[0104] in This is the corresponding degree matrix. The node feature representation of layer l is used to effectively capture the local and global causal dependencies between device parameters through multi-layer graph convolution operations, while introducing an adversarial causal loss function:

[0105] And causal consistency constraints:

[0106] To ensure that the generated samples can pass the discriminator's true / false test and strictly follow the pre-discovered engineering causal logic, GradCAM technology is used to visualize the generator's attention distribution and compare it with the real causal weight matrix for verification. G(z): the image generated by the generator. In addition, an uncertainty-aware generator was also designed.

[0107] and the corresponding uncertainty loss function:

[0108] This allows for the quantification of cognitive and accidental uncertainties during the generation process, thereby providing a credibility assessment for each generated sample and avoiding the generation of engineering-unreasonable combinations of abnormal equipment states.

[0109] Finally, the quality-verified and balanced training set is input into a Transformer-based multi-task deep learning prediction framework for ensemble model training. This prediction framework employs a multi-head self-attention mechanism to capture long-range dependencies between device features.

[0110]

[0111] Furthermore, a causal mask matrix is ​​incorporated into the attention calculation to ensure that the model's prediction process strictly follows temporal causality:

[0112] Simultaneously, a Bayesian deep ensemble method is employed for uncertainty quantification and robust prediction:

[0113]

[0114] A dynamic balance among multiple learning objectives is achieved through an adaptive loss balancing mechanism.

[0115]

[0116]

[0117] in, T To control the smoothness of the weight distribution for temperature parameters, For the task At any momentt In addition to weighting importance, adversarial training stabilization techniques are introduced to enhance the model's robustness to input noise and distribution bias:

[0118]

[0119] Finally, counterfactual reasoning and causal effect quantification are performed using causal effect estimation and dual robust estimators:

[0120]

[0121] in The average treatment effect is used to assess the causal impact of equipment maintenance interventions. The propensity score reflects the probability that a device will receive a specific treatment. The regression function predicts the equipment status under different treatment conditions. This causal reasoning framework can provide a scientific theoretical basis and quantitative effect evaluation for equipment maintenance decisions.

[0122] The entire technical solution combines cutting-edge deep causal inference theory with advanced generative adversarial network technology, and comprehensively utilizes multiple cutting-edge artificial intelligence technologies such as variational Bayesian inference, graph neural networks, meta-learning, tensor decomposition, Transformer architecture, and uncertainty quantification. While thoroughly solving the problem of imbalanced industrial data, it ensures the engineering rationality, physical consistency, and statistical reliability of generated samples. It provides a theoretically rigorous, technologically advanced, practical, and efficient complete solution for equipment failure prediction and industrial artificial intelligence applications, and has broad industrial application prospects and important academic value.

[0123] Example 1—Solving the problem of unbalanced bearing fault data: This invention verifies the performance of bearing failure, a typical equipment fault. The original dataset contains 10,000 equipment operation samples, of which only 100 (1%) are bearing failure samples and 9,900 (99%) are normal operation samples, resulting in an imbalance ratio of 99:1. This extreme imbalance makes it impossible for traditional machine learning algorithms to effectively identify faulty equipment.

[0124] Dataset features: Includes 32 equipment monitoring features such as rotational speed, vibration, temperature, power, load rate, lubricating oil pressure, and running time.

[0125] Specific problems faced: Traditional SMOTE methods may generate samples with engineering inconsistencies, such as "500 rpm vibration frequency of 5000 Hz," ignoring the physical causal chain of "load → power → temperature → bearing wear → fault." During model training, the model tends to predict all samples as normal, resulting in an extremely low fault sample recognition rate.

[0126] Example 2—Missing Value Handling and Causal Relationship Discovery: Step 2.1, Handling Missing Values: The original dataset contained 12% missing values. Analysis revealed: The high-frequency vibration detection feature had a missing rate of 55%, caused by sensor malfunction, and was therefore directly deleted. Basic parameters such as temperature and rotational speed had a missing rate of only 3%, which were filled using conditional mean values ​​based on equipment model and operating conditions. Maintenance record-related features had a missing rate of 18%, which were intelligently filled based on equipment operating patterns.

[0127] IF Equipment Model = Heavy-duty Bearing AND Operating Environment = High Temperature: Missing Temperature Value = Average of this group + Load Correction Factor ELSE: Missing temperature value = average value of the corresponding group Step 2.2, Discovery of Causal Relationships: Key causal relationships were identified through engineering literature and equipment expert knowledge: Load rate → Power consumption → Temperature rise Rotational speed → Vibration frequency → Bearing wear Lubrication conditions → Coefficient of friction → Probability of failure Runtime → Fatigue Level → Fault Severity Construct a causal relationship graph containing 25 nodes and 42 edges.

[0128] Example 3—Causality Quantification and Weight Calculation Step 3.1, Calculation of causal strength: Partial correlation analysis was used to identify the strength of the primary causal relationships: Load factor → Temperature: r = 0.79 (strong correlation) Rotational speed → Vibration: r = 0.84 (extremely strong correlation) Temperature → Failure Risk: r=0.71 (Strong Correlation) Weak relations with |r| < 0.2 were filtered out, and 21 strong causal relations were retained.

[0129] Step 3.2, Data Standardization: Perform Z-score standardization on all numerical features to ensure the comparability of features with different dimensions.

[0130] Example 4—Implementation of Causal Constraint Generative Adversarial Network Step 4.1, Design of the Causal Constraint Generator: Generator architecture: A causal weight matrix is ​​incorporated into the causal attention layer to ensure that the generation process follows engineering causal logic: Input layer (100-dimensional noise) → Fully connected layer (256) → Causal attention layer → Fully connected layer (128) → Output layer (32-dimensional features) Step 4.2, Discriminator Design: The discriminator not only determines whether a sample is genuine or fake, but also assesses the rationality of the project: Input layer (32-dimensional features) → Fully connected layer (128) → Dropout (0.3) → Fully connected layer (64) → Output layer (2-dimensional: true / false + reasonableness) Step 4.3, Causal Consistency Loss Function: The total loss function = adversarial loss + λ × causal consistency loss. Where λ = 0.5, balancing generation quality and causal constraints.

[0131] Step 4.4, Data Integration: 9900 synthetic fault samples were generated and merged with the original 100 fault samples to form 10000 fault samples, which were then combined with 9900 normal samples to form a balanced dataset.

[0132] Example 5—Classification Training and Prediction Using 10-fold cross-validation, we trained SVM and Random Forest classifiers respectively: SVM parameters: RBF kernel, C=1.0, gamma=0.1 Random forest parameters: 100 decision trees, maximum depth = 10 Example 6—Experimental Results and Problem Solving Effects This invention performs weight calculation and quantitative analysis on the causal relationships in the data to obtain a causal weight matrix, with the following preferred scheme: Step 6.1: Calculation of causal relationship strength. Partial correlation analysis can reveal the causal strength between various features. Partial correlation coefficients are used to analyze the causal strength between features and the target variable, filtering out relationships with low causal strength (i.e., |r| < 0.2). The strengths of the main causal relationships obtained after analysis are shown in Table 1.

[0133] Table 1 Strength of Primary Causal Relationships

[0134] As can be seen from Table 1, the causal relationship between workload and equipment pressure is the strongest in equipment failure prediction (r=0.84), which is consistent with common sense in engineering.

[0135] Step 6.2: Test the trained prediction model on the test data. After obtaining the output results, evaluate the results using accuracy, recall, F1 score, precision, and AUC as evaluation criteria. To verify the effectiveness of this invention, experiments were conducted using the following imbalanced dataset: The equipment failure dataset contains industrial equipment operation data, where the failure sample accounts for approximately 1%, and the imbalance ratio is approximately 100:1.

[0136] Experimental results show that the causal inference-based method proposed in this invention significantly outperforms existing baseline methods on all test datasets. In the extremely imbalanced task of device failure prediction, the advantages of this method are evident. Compared to the SMOTE method, the accuracy is improved by 22.6% (from 0.698 to 0.856), the F1 score by 27.5% (from 0.652 to 0.831), and the AUC value by 16.0% (from 0.786 to 0.912). Even compared to the state-of-the-art VAE-Sampler method, the accuracy is still improved by 9.6%, and the F1 score by 11.1%.

[0137] Table 2 Performance Comparison Results

[0138] As can be seen from Table 2, the method of the present invention achieved the best performance on all evaluation metrics. On the extremely imbalanced equipment failure dataset, the performance improvement was significant, with an accuracy improvement of 22.6% and an F1 score improvement of 27.5%, demonstrating the superiority of the present invention in handling extremely imbalanced problems.

[0139] The equipment failure prediction method based on industrial causal logic provided by the present invention, through the above methods, has the following characteristics: 1) Completely resolve the data imbalance problem: Quantification effect: Through generative adversarial networks with causal constraints, high-quality minority class samples can be effectively generated, reducing the imbalance ratio from 100:1 to 1:1 and significantly improving the model training effect.

[0140] Quality assurance: The generated samples not only balance the data distribution in quantity, but more importantly, maintain the same statistical characteristics and engineering rationality as the real samples in quality.

[0141] Highly adaptable: The number of samples generated can be flexibly adjusted according to the specific degree of imbalance, making it suitable for data imbalance scenarios of varying severity.

[0142] 2) Ensure the rationality of the project: Causal consistency: The generated equipment status samples strictly follow the engineering causal logic, avoiding unreasonable engineering samples generated by traditional methods, such as avoiding sample combinations that violate physical common sense, such as "speed 1000rpm but temperature 200℃" or "load rate normal but power consumption extremely high".

[0143] Physical rationality: The complexity of the equipment's physical system was taken into account to ensure that the relationships between various parameters in the generated sample conform to the equipment's operating rules.

[0144] Engineering credibility: After evaluation by equipment experts, the engineering rationality score of the generated sample reached 94.3%, which is significantly higher than the 60% of the traditional SMOTE method.

[0145] 3) Significantly improves prediction accuracy: Improved accuracy: On the bearing fault dataset, the accuracy is improved by 22.6% (from 69.8% to 85.6%) compared with the traditional SMOTE method, meeting the requirements of engineering applications.

[0146] Improved recall rate: The recall rate for a minority of devices increased from 65.2% to 85.4%, meaning that more faulty devices can be identified and the false negative rate is reduced.

[0147] F1 score optimization: The F1 score improved by 27.5% (from 65.2% to 83.1%), achieving a better balance between precision and recall.

[0148] Improved AUC: The AUC value increased by 16.0% (from 78.6% to 91.2%), indicating a significant improvement in the overall classification performance of the model.

[0149] 4) Reduce equipment downtime: Early identification: By improving the accuracy of early identification of equipment faults, engineers can make accurate diagnoses in the early stages of a fault.

[0150] Timely maintenance: Early and accurate diagnosis enables equipment to receive timely maintenance, preventing the malfunction from worsening.

[0151] Expected results: It is expected to increase the early diagnosis rate of equipment failures such as bearing failures from the current 30% to more than 80%, thereby significantly reducing equipment downtime losses caused by false or missed detection of equipment failures.

[0152] 5) Enhance model interpretability: Causal visualization: Through causal relationship diagrams, engineers can intuitively understand the model's prediction logic and the causal relationships between various parameters.

[0153] Engineering explainability: Predictions can be explained using engineering terminology and causal relationships, increasing engineers' trust in the results.

[0154] Decision support: It provides reliable evidence and explanations for maintenance decisions, helping engineers make more accurate diagnoses.

[0155] 6) Improve the generalization ability of the method: Cross-factory applicability: Due to the incorporation of universal engineering causal relationships, this invention maintains stable performance across data from different factories and manufacturing enterprises.

[0156] Cross-device extension: The causal inference framework can be extended to other device failure types, such as motor failures and hydraulic system failures, demonstrating good versatility.

[0157] Stable across datasets: Validated on multiple datasets from different sources, it demonstrates stable performance improvements.

Claims

1. A method for predicting equipment failures based on industrial causal logic, characterized in that, Includes the following steps: Step 1: Obtain the monitoring dataset of unbalanced equipment; Step 2: Handle missing values ​​in the unbalanced equipment monitoring dataset and establish a causal relationship graph; Step 3: Quantify causal relationships and calculate the causal weight matrix; Step 4: Construct a causal constraint generative adversarial network, use the causal weight matrix as attention weights to generate synthetic fault samples, then merge them with the original fault samples to form a balanced fault sample set, and finally merge the balanced fault sample set with the normal sample set to form a balanced dataset. Step 5: Train a classifier based on the balanced dataset and output the fault prediction results.

2. The equipment failure prediction method based on industrial causal logic as described in claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Synchronously collect basic equipment information, real-time operating parameters, multi-dimensional sensor data, historical maintenance records, and fault markers through the industrial data acquisition system; Step 1.2: Unify the fault markers by encoding "fault marker = 0" to indicate normal equipment and "fault marker = 1" to indicate faulty equipment; Step 1.3: Align all collected data according to timestamps to form an unbalanced equipment monitoring dataset as the original fault sample.

3. The equipment failure prediction method based on industrial causal logic as described in claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Identify missing values ​​in the unbalanced equipment monitoring dataset, delete sensor features with a missing ratio ≥ 50%, and fill in features with a missing ratio < 50% using engineering rationality. Step 2.2: Input the dataset after missing values ​​processing into the causal discovery algorithm to identify the causal dependencies between device operating parameters, and output the identification results in the form of a directed acyclic graph to obtain the causal relationship graph.

4. The equipment failure prediction method based on industrial causal logic as described in claim 3, characterized in that, The filling method in step 2.1 is as follows: for continuous operating parameters with a missing ratio of <50%, K-nearest neighbor filling based on similar equipment operating conditions is used for completion. For discrete state features with a missing proportion of less than 50%, rule-based reasoning based on decision trees is used for imputation; for time-series sensor data with a missing proportion of less than 50%, long short-term memory networks are used for sequence prediction imputation; for parameter combinations with physical constraints, constraint optimization is used to ensure that the imputed values ​​conform to engineering practice.

5. The equipment failure prediction method based on industrial causal logic as described in claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Perform partial correlation coefficient analysis, mutual information calculation, Granger causality test and transfer entropy analysis on each pair of variables in the causal relationship diagram in sequence to obtain the strength value of each causal relationship and filter out weak correlation relationships; Step 3.2: Standardize, encode one-hot, and normalize the numerical, categorical, and time-series features respectively; Step 3.3: Use the following fusion formula to fuse the results of multiple methods with the domain expert knowledge weights, and output a causal weight matrix. : in, Indicates the first i The weight matrix obtained by the causal analysis method Indicates the corresponding fusion weights. Represents the knowledge weight of domain experts. β The weighting coefficients representing expert knowledge. k This indicates the number of causal analysis methods.

6. The equipment failure prediction method based on industrial causal logic as described in claim 5, characterized in that, Step 3.1 specifically involves: First, partial correlation coefficients are used to analyze the linear causal strength between operating parameters and fault development. The calculation formula is as follows: in, Represents control variables Z back X and Y The partial correlation coefficient, express X and Y The simple correlation coefficient, Z Indicates control variables; Then, the mutual information method is used to evaluate the strength of nonlinear causality. : in, Representing variables X and Y The mutual information between variables is used to assess the strength of the nonlinear causal relationship between the two variables. express X and Y The joint probability distribution of ; Representing variables X The marginal probability distribution; Representing variables Y The marginal probability distribution; Next, the Granger causality test was used to assess the temporal causality. : in, Indicates from variable X to variable Y The Granger causality test statistic follows a certain order. F Distribution, used for testing X Is it correct? Y It exhibits Granger causality; To represent the sum of squared residuals of the restricted model, use only... Y Historical value prediction Y The future value; Represents the sum of squared residuals of an unrestricted model, and uses X and Y Historical value prediction Y The future value; p Indicates the lag order. n Indicates the total number of samples; Next, the transitive entropy method is used to evaluate the strength of temporal causality. : in, express At any moment of 3D historical state vector express At any moment of 3D historical state vector and These are the embedding dimensions, Represent the corresponding probability density function, through Nearest neighbor method estimation; Finally, relationships with low causal strength, i.e., threshold |r|<0.2, were selected and filtered.

7. The equipment failure prediction method based on industrial causal logic as described in claim 1, characterized in that, Step 4 includes the following steps: Step 4.1: Construct a causal constraint generator. The input layer receives 100-dimensional random noise, which sequentially passes through a 256-dimensional fully connected layer, a causal attention layer, a 128-dimensional fully connected layer, and a 32-dimensional output layer to output a feature vector. The causal attention layer incorporates the causal weight matrix obtained in Step 3 using the following formula. Constrained output feature vector: in, Q , K , V These represent the query, key, and value matrices, respectively. α This represents the causal constraint strength parameter; simultaneously, position encoding is added using the following formula: in, pos For location, i For dimensional indexing, For the embedding dimension of the model; Step 4.2: Construct a multi-task discriminator. The input layer receives a 32-dimensional feature vector, which is then passed through a 128-dimensional fully connected layer, a Dropout layer, a 64-dimensional fully connected layer, and a 2-dimensional output layer to output the results. The discriminator calculates the physical constraint violation loss using the following formula. : in, Indicates the first i A physical constraint function, To constrain tolerance, m To constrain the total number, Tolerance / threshold; and adversarial training with the generator, using the WGAN-GP loss function; Step 4.3: Calculate the adversarial losses separately. Loss of causal consistency Diversity loss and feature matching loss Among them, combating losses The calculation formula is: in, This indicates that the discriminator is effective against real samples. The judgment result, The generator is based on noise. The generated samples, where 𝔼 represents the mathematical expectation, Represents the true data distribution. Indicates noise distribution; Loss of causal consistency The calculation formula is: in, This represents the relevance matrix of the generated samples. A causal weight matrix representing real data. d Indicates feature dimension; Diversity loss Calculated based on sample similarity within a batch: in, B For batch size, , All are generated samples. ε For numerical stability parameters; Feature matching loss The calculation formula is: in, Indicates the first The weighting coefficients of each project constraint. Indicates the first Each engineering constraint function, This represents the state sample of the synthesis device generated by the generator. This represents a sample of real equipment operating data. This represents the total number of engineering constraints. This represents the square of the L2 norm, used to calculate the degree of difference; The total loss is then obtained by weighted summation of the four types of losses using the following formula. Backpropagation updates the generator and discriminator parameters: Among them, the weighting coefficient , , , The constraint strength parameters are determined using a grid search method based on the validation set performance. α Adaptive update via the following formula: in, η For learning rate, For the updated strength parameters, Current constraint strength parameters To The gradient; Step 4.4: After training is completed, generate a number of synthetic fault samples equal to the number of normal samples. Merge the synthetic fault samples with the original fault samples to form a balanced fault sample set. Then merge the balanced fault sample set with the normal sample set to form a balanced dataset. Perform stratified sampling on the balanced dataset to obtain the training set and the test set.

8. The equipment failure prediction method based on industrial causal logic as described in claim 7, characterized in that, Step 5 includes the following steps: Step 5.1: Input the training set into the Support Vector Machine (SVM) for training. The SVM fault diagnosis model is obtained through training using the following formula: in, X Represents equipment characteristic data, W Represents the weight vector. b Indicates the bias term. f ( x Used to determine if the equipment is faulty; Step 5.2: Input the training set into the Random Forest for training, and obtain the RF fault diagnosis model through the following formula: in, I It is an indicator function. The prediction result of the m-th classifier. For real labels, if ,but I =1; otherwise I =0, M Indicates the number of decision trees; Step 5.3: Input the test set into the SVM fault diagnosis model and the RF fault diagnosis model respectively, and output the fault prediction results for each. Step 5.4: Use a voting mechanism to merge the fault prediction results of the SVM fault diagnosis model and the RF fault diagnosis model to obtain the final fault prediction result.

Images