Intelligent multi-sensor fusion fault prediction and diagnosis method and system for depaneling machine

Abstract

The present application discloses an intelligent multi-sensor fusion fault prediction and diagnosis method and system for a depaneling machine, the method comprising: collecting data from a variety of sensors on the depaneling machine; performing random smoothing secondary sampling preprocessing on the sensor data; establishing a causal relationship model using a causal depth set of off-policy evaluation under spatiotemporal interference, and performing off-policy learning on the sensor data; based on the results of the off-policy learning, extracting features of the sensor data and generating causal sequence data; based on the extracted features and the causal sequence data, applying a hybrid model to monitor the state of the depaneling machine in real time, predict potential faults, and generate maintenance suggestions.

Description

Technical Field

[0001] The present application relates to the field of intelligent manufacturing technology, and in particular to an intelligent multi-sensor fusion method and system for predicting and diagnosing faults of a panel splitter. Background Art

[0002] The splitter is an important automated equipment in the electronics manufacturing industry, used to split large printed circuit boards into multiple small circuit boards. As the requirements for precision and efficiency in electronic product manufacturing continue to increase, the performance and stability of the splitter are becoming more and more important. However, the traditional splitter fault diagnosis and prediction methods have the following problems:

[0003] The data from a single sensor is not sufficient to fully reflect the operating status of the depaneling machine, resulting in low accuracy and reliability of fault diagnosis.

[0004] Traditional data processing methods have difficulty in effectively handling noise and interference in multi-sensor data, which affects the accuracy of fault prediction.

[0005] Traditional methods make it difficult to monitor the status of the depaneling machine in real time and predict potential failures in a timely manner, and are unable to provide timely and effective suggestions for equipment maintenance. Summary of the invention

[0006] In view of this, the present application provides an intelligent multi-sensor fusion board splitter fault prediction and diagnosis method and system, which solves the problems of insufficient single sensor data, poor data processing effect and insufficient real-time monitoring and prediction capabilities in the prior art.

[0007] An embodiment of the present application provides an intelligent multi-sensor fusion method for predicting and diagnosing faults of a board splitter, comprising: collecting data from multiple sensors on the board splitter; performing random smoothing secondary sampling preprocessing on the sensor data; establishing a causal relationship model using a causal depth set of off-policy evaluation under spatiotemporal interference, and performing off-policy learning on the sensor data; based on the results of the off-policy learning, extracting features of the sensor data and generating causal sequence data; based on the extracted features and the causal sequence data, applying a hybrid model to monitor the status of the board splitter in real time, predict potential faults, and generate maintenance recommendations.

[0008] The random smoothing secondary sampling preprocessing of the sensor data includes: performing initial sampling on the sensor data; performing random smoothing processing on the initial sampling data; and performing secondary sampling on the smoothed initial sampling data.

[0009] The random smoothing process for the initial sampling data includes: defining a smoothing function; generating a random disturbance for each data point in the initial sampling data; and applying the smoothing function to process the data after the random disturbance.

[0010] The method of establishing a causal relationship model using a causal depth set of off-policy evaluation under spatiotemporal interference includes: identifying key variables and establishing causal relationships between variables; performing off-policy learning to learn the optimal strategy from historical data; performing deep ensemble learning to train and integrate multiple deep learning models based on the optimal strategy; processing the spatiotemporal interference of the multiple deep learning models, identifying and processing spatial and temporal interference factors; and constructing a causal relationship model based on the multiple deep learning models after processing.

[0011] Off-policy learning is performed, including: collecting historical data, wherein the historical data includes historical sensor data, maintenance records and fault logs; defining processing procedures and processing results, wherein the processing procedures include maintenance operations and parameter adjustments, and the processing results include equipment performance indicators and fault states; estimating propensity scores, including: using logistic regression or random forest to calculate the probability of taking any processing procedure under given covariates; building a counterfactual regression model to predict the results under different processing procedures; using the inverse probability weighting method to calculate the weighted average treatment effect; applying a double robust method, combining the regression model and the inverse probability weighting to estimate the causal effect, the causal effect includes calculating the average treatment effect and the conditional average treatment effect; using the bootstrap method to estimate the confidence interval of the causal effect; then generating causal sequence data based on the results of the off-policy learning, including: identifying key causal chains, building temporal causal relationships based on the estimated causal effects, and generating sequence data containing causal relationship strength and temporal information.

[0012] The hybrid model includes a hybrid convolutional neural network CNN-Transformer and a state-space model. Before applying the hybrid model to monitor the state of the board separator in real time, the method also includes: constructing a physical model to analyze the physical characteristics of the board separator; using CNN to extract spatiotemporal features from the sensor data; applying the Transformer architecture to perform sequence modeling on the extracted spatiotemporal features to capture long-term dependencies; integrating the state-space model and combining it with the constraints of the physical model to fuse and optimize the hybrid model.

[0013] Utilizing CNN to extract spatiotemporal features from the sensor data includes: extracting local features of the sensor data through multiple convolutional layers; using pooling layers to reduce the dimensions of the local features; applying normalization layers to accelerate training and improve the stability of the CNN model; and integrating feature information through fully connected layers.

[0014] The Transformer architecture is used to perform sequence modeling on the extracted spatiotemporal features, including: converting the spatiotemporal features into a format corresponding to the Transformer through an input embedding layer to form a spatiotemporal feature sequence; applying a multi-head self-attention mechanism to capture the dependency of the spatiotemporal feature sequence; using a feedforward neural network to enhance the nonlinear expression ability of the Transformer; adding position encoding to provide position information of the spatiotemporal feature sequence.

[0015] The integrated state space model, combined with the constraints of the physical model, includes: defining key state variables, including spindle angular velocity, temperature, vibration amplitude and tool wear; establishing state transfer equations and observation equations, wherein the state transfer equations describe the evolution of state variables over time, and the observation equations describe the relationship between state variables and observation data; converting the dynamic equations in the physical model into state space representation, including: discretizing continuous-time dynamic equations; mapping physical parameters to parameters of the state space model; designing regularization terms based on physical constraints to ensure that predictions conform to physical laws; implementing a hybrid filtering algorithm that combines Transformer output and physical model predictions to obtain state estimates and predictions, including: using an extended Kalman filter or an unscented Kalman filter as a filter; inputting the output of the Transformer as an observation value into the filter; using the physical model prediction as prior information; and obtaining the optimal state estimate and prediction through a filtering algorithm.

[0016] The embodiment of the present application also provides an intelligent multi-sensor fusion board splitter fault prediction and diagnosis system, including: an acquisition module, used to collect data from various sensors on the board splitter; a preprocessing module, used to perform random smoothing secondary sampling preprocessing on the sensor data; an off-policy learning module, used to establish a causal relationship model using the causal depth set of off-policy evaluation under spatiotemporal interference, and perform off-policy learning on the sensor data; a generation module, used to extract features of the sensor data and generate causal sequence data based on the results of the off-policy learning; a monitoring module, used to apply a hybrid model to monitor the status of the board splitter in real time based on the extracted features and the causal sequence data, and predict potential faults and generate maintenance suggestions.

[0017] An embodiment of the present application also provides a computer device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the steps of the above-mentioned intelligent multi-sensor fusion board splitter fault prediction and diagnosis method.

[0018] An embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the steps of the above-mentioned intelligent multi-sensor fusion board splitter fault prediction and diagnosis method.

[0019] An embodiment of the present application also provides a computer program product, including computer instructions, characterized in that when the computer instructions are executed by a processor, the steps of the above-mentioned intelligent multi-sensor fusion board splitter fault prediction and diagnosis method are implemented.

[0020] This application has the following technical effects:

[0021] Through multi-sensor data fusion, the operating status of the board splitter is fully captured, improving the accuracy and reliability of fault diagnosis.

[0022] The random smoothing secondary sampling preprocessing method is used to effectively reduce the noise and interference in the sensor data and improve the data quality.

[0023] The causal relationship model is established by using the causal depth set of off-policy evaluation under spatiotemporal interference, which fully explores the causal relationship between sensor data and improves the interpretability and predictive ability of the model.

[0024] Through off-policy learning, historical data is effectively used to learn the optimal maintenance strategy, which improves the accuracy of fault prediction and maintenance decision-making.

[0025] The hybrid model is used to monitor the status of the depaneling machine in real time, predict potential failures and generate maintenance suggestions, thus realizing intelligent management and preventive maintenance of the depaneling machine.

[0026] Combining physical models and data-driven models improves the accuracy and interpretability of fault prediction, making the prediction results more reliable and easier to understand. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0028] Figure 1 A flow chart of a fault prediction and diagnosis method for a panel splitter using intelligent multi-sensor fusion provided in an embodiment of the present application;

[0029] Figure 2 A module structure diagram of a fault prediction and diagnosis method system for a panel splitter using intelligent multi-sensor fusion provided in an embodiment of the present application;

[0030] Figure 3 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0032] Figure 1 FIG. 1 is a flow chart of a method for fault prediction and diagnosis of a panel splitter using intelligent multi-sensor fusion in an embodiment of the present invention. Figure 1 As shown, the embodiment of the present invention includes the following steps:

[0033] Combine the following Figure 1 and Figure 2 , the intelligent multi-sensor fusion board splitter fault prediction and diagnosis method provided in the embodiment of the present application is described in detail.

[0034] S1: Collect data from various sensors on the depaneling machine.

[0035] In this step, the system collects operating data through various sensors installed on the depaneling machine. These sensors include but are not limited to temperature sensors, vibration sensors, current sensors, sound sensors, etc. Each sensor can capture different aspects of the depaneling machine's operating status and provide comprehensive data support for subsequent fault prediction and diagnosis.

[0036] For example, temperature sensors can monitor temperature changes in key components such as spindles and motors; vibration sensors can detect the vibration of equipment and reflect the wear status of mechanical parts; current sensors can monitor the working status of motors and other electrical components; sound sensors can capture abnormal noise and detect potential problems early.

[0037] S2: Subsampling preprocessing of sensor data with random smoothing.

[0038] S2.1: Perform initial sampling on the sensor data.

[0039] Initial sampling is to extract representative samples from the raw sensor data. The selection of sampling frequency needs to be determined according to the working characteristics of the depaneling machine and the characteristics of different sensors. For example, for fast-changing parameters (such as vibration), a higher sampling frequency is required; while for slowly changing parameters (such as temperature), the sampling frequency can be relatively low.

[0040] S2.2: Perform random smoothing processing on the initial sampling data.

[0041] S2.2.1: Define smoothing function.

[0042] The selection of smoothing function is crucial to the effect of data preprocessing. Commonly used smoothing functions include moving average, exponential smoothing, Gaussian smoothing, etc. In this embodiment, a smoothing function suitable for processing time series data can be selected, such as exponentially weighted moving average (EWMA).

[0043] The EWMA function can be expressed as:

[0044] S t = α* Y t + (1 -α) * S t-1

[0045] Among them, S t is the smoothed value at time t, Y t is the observed value at time t, and α is the smoothing factor (0 < α < 1).

[0046] S2.2.2: Generate a random perturbation for each data point in the initial sampled data.

[0047] Generating random perturbations for each data point can help the model better cope with noise and fluctuations in the real environment. The magnitude of the random perturbations should be determined based on the characteristics of the sensor and the measurement accuracy to ensure that no excessive errors are introduced.

[0048] For example, random perturbations can be generated using a normal distribution:

[0049] Perturbation = N(0, σ 2 )

[0050] Among them, σ 2 is the variance determined by the sensor accuracy.

[0051] S2.2.3: Apply the smoothing function to the randomly perturbed data.

[0052] Applying the defined smoothing function to the data with added random perturbations can effectively reduce the impact of noise while retaining the important features of the data.

[0053] In step S2.2.3, the present invention applies a smoothing function to the data to which random disturbances are added. The purpose of this process is to reduce the impact of noise while retaining important features of the data. Let the present invention take the data of the temperature sensor as an example to illustrate this process.

[0054] Example: An embodiment of the present invention has a temperature sensor that collects data once a second within 10 seconds.

[0055] The original data is as follows (unit: Celsius):

[0056] [25.0, 25.2, 25.1, 25.3, 25.2, 25.4, 25.3, 25.5, 25.4, 25.6]

[0057] Step 1: Add random perturbations

[0058] First, the embodiment of the present invention adds a small random disturbance to each data point. The embodiment of the present invention uses a normal distribution with a mean of 0 and a standard deviation of 0.1 to generate the disturbance.

[0059] The data after adding disturbance is as follows:

[0060] [25.05, 25.15, 25.18, 25.22, 25.31, 25.35, 25.25, 25.58, 25.32,25.67]

[0061] Step 2: Apply the smoothing function

[0062] Now, the embodiment of the present invention will use the previously defined exponentially weighted moving average (EWMA) function to smooth these data. The formula of the EWMA function is:

[0063] St = α * Yt + (1 - α) * St-1

[0064] Where St is the smoothed value at time t, Yt is the observed value at time t, and α is the smoothing factor (0 < α < 1).

[0065] Let α = 0.3 in the embodiment of the present invention. The embodiment of the present invention will gradually calculate the smoothed value:

[0066] 1. S1 = 25.05 (initial value is set to the first observation)

[0067] 2. S2 = 0.3 * 25.15 + 0.7 * 25.05 = 25.08

[0068] 3. S3 = 0.3 * 25.18 + 0.7 * 25.08 = 25.11

[0069] 4. S4 = 0.3 * 25.22 + 0.7 * 25.11 = 25.14

[0070] 5. S5 = 0.3 * 25.31 + 0.7 * 25.14 = 25.19

[0071] 6. S6 = 0.3 * 25.35 + 0.7 * 25.19 = 25.24

[0072] 7. S7 = 0.3 * 25.25 + 0.7 * 25.24 = 25.24

[0073] 8. S8 = 0.3 * 25.58 + 0.7 * 25.24 = 25.34

[0074] 9. S9 = 0.3 * 25.32 + 0.7 * 25.34 = 25.33

[0075] 10. S10 = 0.3 * 25.67 + 0.7 * 25.33 = 25.43

[0076] The data sequence after smoothing is:

[0077] [25.05, 25.08, 25.11, 25.14, 25.19, 25.24, 25.24, 25.34, 25.33,25.43]

[0078] This example shows how to apply the EWMA smoothing function to process data with random disturbances added. It can be seen from the embodiment of the present invention that the smoothed data reduces short-term fluctuations but still retains the overall upward trend. This processing method has the following advantages:

[0079] 1. Reduce the impact of random noise: the smoothed data curve is smoother than the original data.

[0080] 2. The overall trend of the data is retained: Although short-term fluctuations are weakened, the overall rising trend of temperature is still obvious.

[0081] 3. Strong adaptability: By adjusting the α value, the embodiment of the present invention can control the degree of smoothing. A smaller α value will produce a smoother result, while a larger α value will respond to data changes more quickly.

[0082] 4. High computational efficiency: EWMA only needs to save the smoothed value of the previous moment, so it has low memory requirements and fast computation speed, making it suitable for real-time processing of large amounts of sensor data.

[0083] This random smoothing secondary sampling preprocessing method can effectively improve the data quality of subsequent analysis and model training, thereby improving the accuracy of fault prediction and diagnosis of the depaneling machine.

[0084] S2.3: Perform secondary sampling on the smoothed initial sampling data.

[0085] The purpose of subsampling is to further reduce the amount of data and improve processing efficiency while retaining key information. The frequency of subsampling can be determined according to the needs of subsequent analysis and is usually lower than the initial sampling frequency.

[0086] S3: Building a causal relationship model using the causal depth set of off-policy evaluation under spatiotemporal interference, and performing off-policy learning of the sensor data.

[0087] S3.1: Identify key variables and establish causal relationships between them.

[0088] In this step, it is necessary to identify the key variables that have a significant impact on the performance and fault diagnosis of the depaneling machine based on professional knowledge and data analysis results. These variables include spindle speed, cutting force, vibration amplitude, temperature, etc. Then, the causal relationship between these variables is preliminarily established through causal inference methods (such as structural equation models or Bayesian networks).

[0089] S3.2: Perform off-policy learning to learn the optimal strategy from historical data.

[0090] S3.2.1: Collect historical data, wherein the historical data includes historical sensor data, maintenance records, and fault logs.

[0091] This step requires the integration of multiple data sources, including historical sensor data, equipment maintenance records, and fault logs. These data should cover a long enough time span to capture the performance of the depaneling machine under different working conditions and different maintenance strategies.

[0092] S3.2.2: Define the processing process and processing results, where the processing process includes maintenance operations and parameter adjustments, and the processing results include equipment performance indicators and fault conditions.

[0093] Clearly define what is a "process" (such as a specific maintenance operation or parameter adjustment) and what is a "process result" (such as a performance indicator or fault status of the equipment). This lays the foundation for subsequent causal inference.

[0094] S3.2.3: Estimate propensity scores, including using logistic regression or random forests to calculate the probability of taking either treatment given the covariates.

[0095] Propensity score estimation is to balance the distribution of covariates between different treatment groups. Propensity scores can be estimated using logistic regression models (for binary treatments) or multinomial logistic regression (for multi-class treatments). Random forests are also a powerful alternative, especially when there are complex nonlinear relationships between covariates.

[0096] Specifically, the propensity score refers to the probability that an individual will accept a specific treatment given a set of covariates. In the context of fault prediction and diagnosis of depaneling machines, "treatment" refers to a specific maintenance operation or parameter adjustment, while covariates include equipment operating parameters, environmental conditions, etc.

[0097] The main purposes of estimating propensity scores are to:

[0098] 1. Balance the distribution of covariates between the treatment and control groups;

[0099] 2. Reduce the confounding effect caused by selection bias;

[0100] 3. Provide a basis for subsequent causal inference;

[0101] Here is a detailed description of how to estimate propensity scores using logistic regression and random forests:

[0102] 1. Use logistic regression to estimate propensity scores;

[0103] Logistic regression is a classic method for estimating propensity scores, especially suitable for cases where the variables are binary.

[0104] step:

[0105] a) Prepare the data: Use the covariates as independent variables and the treatment indicator (0 or 1) as the dependent variable.

[0106] b) Building a logistic regression model:

[0107] P(T=1|X) = 1 / (1 + e βX )

[0108] where T is the treatment indicator, X is the vector of covariates, and β is the vector of parameters to be estimated.

[0109] c) Estimate the model parameters using maximum likelihood estimation.

[0110] d) For each sample, use the trained model to predict the probability of receiving treatment, which is the propensity score.

[0111] The embodiment of the present invention has the following covariates: spindle operation time (x1), ambient temperature (x2), and time from the last maintenance to the present (x3). Process T indicates whether a specific maintenance operation has been performed.

[0112] The logistic regression model is as follows:

[0113] log(P(T=1) / (1-P(T=1))) = β0 + β1x1 + β2x2 + β3x3

[0114] After the model is trained, the embodiment of the present invention can calculate P(T=1), ie, the propensity score, for each sample.

[0115] 2. Estimating Propensity Scores Using Random Forests

[0116] Random forest is a more flexible method that can capture nonlinear relationships and interactions between covariates.

[0117] step:

[0118] a) Prepare data: Same as logistic regression.

[0119] b) Build a random forest model:

[0120] Sampling with replacement from the original training data creates multiple subsamples.

[0121] A decision tree is trained for each subsample, but only a subset of randomly selected features are considered at each split.

[0122] Repeat the above process to generate multiple decision trees (for example, 500 trees).

[0123] c) For each sample, use the average of the predicted probabilities of all decision trees as the final propensity score.

[0124] Using the same covariates (x1, x2, x3) as above, random forests can automatically capture the nonlinear relationships between them. For example, there is an interaction between spindle running time (x1) and ambient temperature (x2) that affects maintenance decisions. Random forests are able to automatically model this complex relationship.

[0125] 3. Model selection and diagnostics

[0126] The choice between logistic regression and random forest can be based on the following considerations:

[0127] If the relationship between the covariates and the treatment is simple, logistic regression may be chosen.

[0128] If there are complex nonlinear relationships or high-order interactions, random forests are more appropriate.

[0129] Whichever method you choose, you should perform model diagnostics:

[0130] a) Use cross-validation to evaluate the predictive performance of the model.

[0131] b) Check the distribution of propensity scores to ensure that there is sufficient overlap between the treatment and control groups.

[0132] c) Assess balance: Check whether the distribution of covariates in the treatment and control groups is similar in different propensity score strata.

[0133] 4. Handling multiple types of processing

[0134] If there are multiple processing options (for example, different types of maintenance operations), you can:

[0135] Multinomial logistic regression was used;

[0136] Build multiple binary classifiers (one-vs-the-rest method);

[0137] Directly use random forests that can handle multi-class problems;

[0138] 5. Using Estimated Propensity Scores

[0139] Once the propensity scores are estimated, they can be used to:

[0140] Stratified analysis: The sample is divided into several strata based on the propensity score, and the treatment effects are compared within each stratum.

[0141] Matching: For each treated sample, find an untreated sample with a similar propensity score.

[0142] Weighted: Use the inverse of the propensity score as the weight to balance the treatment and control groups.

[0143] Through this method, the embodiment of the present invention can more accurately estimate the causal effects of different maintenance strategies or parameter adjustments on the performance and failure probability of the board splitter, thereby providing a scientific basis for formulating the optimal maintenance strategy.

[0144] S3.2.4: Construct counterfactual regression models to predict outcomes under different treatments.

[0145] Counterfactual regression models are used to predict what would happen if different treatments were taken given the covariates. This can be done by building multiple regression models, one for each treatment.

[0146] The purpose of the counterfactual regression model is to predict what results would be produced if the embodiment of the present invention took a different treatment on a sample than was actually taken. In the context of fault prediction and diagnosis of a depaneling machine, this means that the embodiment of the present invention can predict how the performance indicators of the depaneling machine would change if different maintenance operations or parameter adjustments were taken.

[0147] Detailed explanation of the process of building a counterfactual regression model:

[0148] 1. Data Preparation

[0149] First, the embodiment of the present invention needs to prepare data for training the counterfactual model. The data should include:

[0150] Covariates (X): such as the operating parameters of the board splitter, environmental conditions, etc.

[0151] Processing indicator (T): indicates the actual maintenance operation or parameter adjustment taken;

[0152] Observed results (Y): such as equipment performance indicators, failure rates, etc.

[0153] 2. Model selection

[0154] Counterfactual regression models can take many forms, common ones include:

[0155] a) Linear regression model:

[0156] Y = β0 + β1X + β2T + β3(X*T) + ε

[0157] where X*T represents the interaction term between the covariate and the treatment.

[0158] b) Generalized linear models: suitable for outcome variables with non-normal distribution, such as Poisson regression (for count data) or logistic regression (for binary outcomes).

[0159] c) Non-parametric models: such as random forests and gradient boosted trees, which can capture complex nonlinear relationships.

[0160] d) Neural Networks: They excel in processing high-dimensional data.

[0161] 3. Model Training

[0162] Taking the linear regression model as an example, the training process is as follows:

[0163] a) For each treatment T, train a separate model:

[0164] YT = βT0 + βT1X + εT

[0165] b) Train these models using the observed data. For example, for samples that actually received the treatment, use the model with T=1; for samples that did not receive the treatment, use the model with T=0.

[0166] c) Estimate the model parameters using appropriate methods (e.g., least squares method).

[0167] 4. Counterfactual Predictions

[0168] After the model is trained, the embodiment of the present invention can perform counterfactual predictions:

[0169] a) For each sample, regardless of the treatment it actually received, predictions are made using the models corresponding to all treatments.

[0170] b) For example, for a sample that actually received treatment (T=1):

[0171] Use the T=1 model to predict actual results;

[0172] Use the T=0 model to predict what would happen if no treatment was received;

[0173] Example: Take a specific board splitter maintenance scenario as an example:

[0174] Covariate X:

[0175] x1: spindle running time (hours);

[0176] x2: time from the last maintenance to now (days);

[0177] x3: average cutting speed (m / min);

[0178] Processing T:

[0179] T=0: no maintenance;

[0180] T=1: Perform routine maintenance;

[0181] Result Y: Failure rate in the next cycle (times / day);

[0182] The embodiment of the present invention will obtain two models:

[0183] For T=0 (no maintenance):

[0184] Y0 = 0.1 + 0.005x1 + 0.01x2 + 0.02x3

[0185] For T=1 (maintenance):

[0186] Y1 = 0.05 + 0.002x1 + 0.005x2 + 0.01x3

[0187] Now, there is a sample of the embodiment of the present invention:

[0188] x1 = 1000, x2 = 30, x3 = 5

[0189] Counterfactual predictions:

[0190] If no maintenance (T=0): Y0 = 0.1 + 0.005*1000 + 0.01*30 + 0.02*5 = 5.5 times / day

[0191] If maintenance (T=1): Y1 = 0.05 + 0.002*1000 + 0.005*30 + 0.01*5 = 2.35 times / day

[0192] This prediction suggests that maintenance can significantly reduce the failure rate.

[0193] By constructing and applying a counterfactual regression model, the embodiments of the present invention can more scientifically and accurately predict the impact of different maintenance strategies on the performance of the panel splitter, thereby supporting more informed decision-making and improving the reliability and production efficiency of the equipment.

[0194] S3.2.5: Calculate the weighted average treatment effect using the inverse probability weighting method.

[0195] Inverse probability weighting balances differences between treatment groups by assigning a weight to each sample. The weight is usually defined as the inverse of the observed probability of the treatment. This approach helps reduce the effects of selection bias.

[0196] Inverse probability weighting (IPW) is a commonly used causal inference method that estimates the average treatment effect by weighting the samples to balance the differences between the treatment and control groups.

[0197] The detailed steps are as follows:

[0198] 1. Calculate weights:

[0199] For each sample i, the weight is calculated based on the treatment Ti it actually receives and the estimated propensity score e(Xi):

[0200] Wi = Ti / e(Xi) + (1 - Ti) / (1 - e(Xi))

[0201] Where Ti is the treatment indicator (1 means receiving the treatment and 0 means not receiving the treatment), and e(Xi) is the propensity score of sample i.

[0202] 2. Calculate the weighted average treatment effect:

[0203] ATE_IPW = (1 / N) * Σ[Wi * Yi * (Ti - e(Xi))] / [(1 / N) * Σ(Ti - e(Xi))²]

[0204] Where N is the total number of samples and Yi is the observed outcome.

[0205] 3. Calculate the standard error:

[0206] The standard error of ATE_IPW can be calculated using the Bootstrap method or the Sandwich estimator.

[0207] 4. Construct confidence intervals:

[0208] A 95% confidence interval was constructed based on the estimated ATE_IPW and its standard error.

[0209] Example:

[0210] The embodiment of the present invention has the following data, as shown in Table 1:

[0211] Table 1

[0212] sample X (covariate) T (Treatment) Y (result) e(X) (propensity score) 1 0.3 1 5 0.6 2 0.7 0 3 0.7 3 0.5 1 6 0.5

[0213] Calculate weights:

[0214] W1 = 1 / 0.6 = 1.67

[0215] W2 = 1 / (1 - 0.7) = 3.33

[0216] W3 = 1 / 0.5 = 2

[0217] Then calculate ATE_IPW (to simplify the calculation process).

[0218] S3.2.6: Apply a doubly robust approach combining regression models and inverse probability weighting to estimate causal effects

[0219] The doubly robust (DR) method combines the advantages of regression models and IPW to provide a more reliable estimate of causal effects. Its advantage is that even if one of the regression model or the propensity score model is misspecified, the estimation results remain consistent.

[0220] The detailed steps are as follows:

[0221] 1. Build a regression model:

[0222] Regression models were constructed for the treatment group and the control group respectively:

[0223] m1(X) = E[Y|X, T=1]

[0224] m0(X) = E[Y|X, T=0]

[0225] 2. Calculate IPW weight:

[0226] Same as step S3.2.5.

[0227] 3. Calculate the doubly robust estimate:

[0228] ATE_DR = (1 / N) * Σ{[Ti * Yi / e(Xi) - (Ti - e(Xi)) * m1(Xi) / e(Xi)]-

[0229] [(1 - Ti) * Yi / (1 - e(Xi)) + (Ti - e(Xi)) * m0(Xi) / (1 - e(Xi))]}

[0230] Among them, m1(Xi) and m0(Xi) are the predicted values ​​of the regression model for sample i when it receives treatment and when it does not receive treatment.

[0231] 4. Calculate the conditional average treatment effect (CATE):

[0232] CATE(x) = m1(x) - m0(x)

[0233] This allows embodiments of the invention to estimate how treatment effects vary with covariates.

[0234] 5. Calculate the standard error and confidence interval:

[0235] You can also use the Bootstrap method.

[0236] Example:

[0237] Continuing to use the above data, the embodiment of the present invention has the following regression model prediction:

[0238] Table 2

[0239] sample m1(X) m0(X) 1 5.2 3.8 2 5.5 3.5 3 5.8 4.2

[0240] The embodiment of the present invention may use these values ​​to calculate ATE_DR (simplifying the calculation process).

[0241] application:

[0242] 1. Heterogeneous treatment effects:

[0243] By analyzing CATE, the embodiment of the present invention can understand how the processing effect varies with the characteristics of the panel splitter (such as usage time, workload).

[0244] 2. Personalized maintenance strategy:

[0245] Based on CATE, the embodiment of the present invention can formulate a personalized maintenance plan for each panel splitter to optimize resource allocation.

[0246] 3. Policy evaluation:

[0247] Use ATE_DR to evaluate the overall effect of different maintenance policies and support decision making.

[0248] 4. Sensitivity analysis:

[0249] The robustness of the estimates is assessed by comparing the results of the IPW and DR methods.

[0250] 5. Exploration of causal mechanisms:

[0251] Analyze which covariates have the greatest impact on the treatment effect and gain a deeper understanding of the mechanism of action of the maintenance operation.

[0252] Through these steps, the embodiment of the present invention can not only estimate the average effect of the maintenance operation, but also understand how the effect varies with the characteristics of the board splitter, thereby supporting more refined maintenance decisions and resource allocation. This method combines the advantages of statistical learning and causal inference, and provides a solid theoretical foundation for the intelligent management of the board splitter.

[0253] S3.2.7: Use bootstrapping to estimate confidence intervals for causal effects.

[0254] Bootstrapping constructs an empirical distribution of causal effect estimates by repeatedly sampling with replacement from the original data and reestimating the causal effect on each sample. This approach can provide confidence intervals for the causal effect estimate, which helps assess the uncertainty of the estimate.

[0255] The confidence interval of the causal effect is estimated using the bootstrap method. Bootstrap is a powerful statistical technique used to estimate the uncertainty of a statistic, and is particularly suitable for complex estimation procedures, such as the method used in causal inference in the embodiments of the present invention.

[0256] Below are the detailed steps to estimate the confidence interval of the causal effect using the bootstrap method:

[0257] 1. Basic principles

[0258] The core idea of ​​the bootstrap method is to create multiple simulated data sets by repeatedly sampling with replacement from the original data, and then recalculate the statistic of interest (in the case of the embodiment of the present invention, the causal effect estimate) on each simulated data set to obtain the distribution of the statistic.

[0259] 2. Detailed steps

[0260] a) Prepare the original dataset;

[0261] In the embodiment of the present invention, there is an original data set D containing N samples, each sample containing a covariate X, a process T and a result Y.

[0262] b) Set the number of repetitions B

[0263] Usually B is set to 1000 or more to ensure stable results.

[0264] c) conduct bootstrap sampling and estimation;

[0265] For i = 1 to B:

[0266] N samples are extracted from the original data set D with replacement to form a new data set Di.

[0267] Propensity scores were reestimated using Di.

[0268] Recalculate causal effects (such as ATE_IPW or ATE_DR) on Di.

[0269] Record this estimate as θi.

[0270] d) Calculate confidence intervals;

[0271] Sort the B θi values ​​from small to large.

[0272] For a 95% confidence interval, find the 25th (2.5th percentile) and 975th (97.5th percentile) values ​​of θi.

[0273] These two values ​​constitute the 95% confidence interval for the estimate of the causal effect.

[0274] 3. Specific implementation examples

[0275] The present embodiment is estimating the average process effect (ATE) of maintenance operations on the failure rate of a depaneler.

[0276] Python code example:

[0277] ```Python

[0278] import numpy as np

[0279] from sklearn.utils import resample

[0280] def estimate_ate(data):

[0281] # This should include the ATE estimation method discussed before in the embodiment of the present invention

[0282] # For simplicity, this function in the embodiment of the present invention directly returns the ATE estimate

[0283] return np.mean(data['Y'][data['T']==1]) - np.mean(data['Y'][data['T']==0])

[0284] # Raw data

[0285] data = {

[0286] 'X': [...], # covariates

[0287] 'T': [...], # Processing indicator

[0288] 'Y': [...] # result

[0289] }

[0290] n_bootstrap = 1000

[0291] ate_bootstrap = np.zeros(n_bootstrap)

[0292] for i in range(n_bootstrap):

[0293] # Sampling with replacement

[0294] boot_data = resample(data, n_samples=len(data['X']))

[0295] # Estimate ATE on bootstrap samples

[0296] ate_bootstrap[i] = estimate_ate(boot_data)

[0297] # Calculate 95% confidence interval

[0298] ci_lower, ci_upper = np.percentile(ate_bootstrap, [2.5, 97.5])

[0299] print(f"95% CI for ATE: ({ci_lower:.2f}, {ci_upper:.2f})")

[0300] ```

[0301] 4. Interpret the results

[0302] The results obtained by the embodiment of the present invention are:

[0303] ```

[0304] 95% CI for ATE: (-0.15, -0.05)

[0305] ```

[0306] This means that the present invention embodiment has 95% confidence that the true average treatment effect of the maintenance operation on the failure rate is between -0.15 and -0.05. A negative value indicates that the maintenance operation reduces the failure rate.

[0307] 5. Application in maintenance decision of PCB splitter

[0308] Risk Assessment: Confidence intervals can help assess the uncertainty of maintenance decisions. For example, if the confidence interval crosses zero, the results need to be interpreted more cautiously.

[0309] Sample size planning: If the confidence interval is too wide, it indicates that more data needs to be collected to obtain a more precise estimate.

[0310] Comparing different maintenance strategies: By comparing the confidence intervals of different maintenance strategies, you can assess which strategy is more likely to produce significant effects.

[0311] Decision support: Combined with cost-benefit analysis, confidence intervals can help managers weigh the potential benefits and risks of maintenance decisions.

[0312] By using the bootstrap method to estimate the confidence interval of the causal effect, the embodiments of the present invention not only obtain a point estimate, but also obtain important information about the uncertainty of this estimate. This is crucial for making robust decisions in real industrial environments because it allows the embodiments of the present invention to evaluate the expected effects of maintenance strategies while considering potential risks.

[0313] S3.3: Perform deep ensemble learning to train and integrate multiple deep learning models based on the optimal strategy.

[0314] Deep ensemble learning improves the generalization and robustness of the model by combining the advantages of multiple deep learning models. In this step, different types of deep learning models (such as convolutional neural networks, recurrent neural networks, long short-term memory networks, etc.) can be selected for training, and each model is optimized based on the optimal strategy learned previously. The outputs of these models are then combined through ensemble methods (such as voting, averaging, or more complex strategies).

[0315] In the context of fault prediction and diagnosis of a depaneling machine, the embodiment of the present invention can use deep ensemble learning to improve the accuracy and robustness of prediction. Let the embodiment of the present invention explain this process in detail through an example.

[0316] Example: Predicting the remaining useful life (RUL) of a panel splitter

[0317] 1. Data Preparation

[0318] The embodiment of the present invention has the following data:

[0319] Input features(X):

[0320] Spindle speed (rpm)

[0321] Cutting force(N)

[0322] Vibration amplitude (mm)

[0323] Temperature (°C)

[0324] Cumulative running time (hours)

[0325] Output(Y):

[0326] Remaining service life (hours)

[0327] 2. Choose the model architecture based on the optimal strategy

[0328] Based on the previous off-policy learning results, the embodiment of the present invention determines the following optimal strategy:

[0329] Use time series data;

[0330] Consider long-term and short-term dependencies;

[0331] Focus on multi-scale features;

[0332] Based on these strategies, the embodiment of the present invention selects the following three deep learning models:

[0333] a) Long Short-Term Memory (LSTM)

[0334] b) One-dimensional Convolutional Neural Network (1D CNN)

[0335] c) Temporal Convolutional Network (TCN)

[0336] 3. Model architecture design

[0337] a) LSTM model:

[0338] ```Python

[0339] class LSTMModel(nn.Module):

[0340] def __init__(self, input_dim, hidden_dim, num_layers, output_dim):

[0341] super(LSTMModel, self).__init__()

[0342] self.lstm = nn.LSTM(input_dim, hidden_dim, num_layers, batch_first=True)

[0343] self.fc = nn.Linear(hidden_dim, output_dim)

[0344] def forward(self, x):

[0345] _, (hidden, _) = self.lstm(x)

[0346] out = self.fc(hidden[-1])

[0347] return out

[0348] ```

[0349] b) 1D CNN Model:

[0350] ```Python

[0351] class CNNModel(nn.Module):

[0352] def __init__(self, input_dim, num_filters, kernel_sizes, output_dim):

[0353] super(CNNModel, self).__init__()

[0354] self.convs = nn.ModuleList(

[0355] nn.Conv1d(input_dim, num_filters, k) for k in kernel_sizes )

[0357] self.fc = nn.Linear(len(kernel_sizes) * num_filters, output_dim)

[0358] def forward(self, x):

[0359] x = x.transpose(1, 2) # (batch, input_dim, time_steps)

[0360] conv_outputs = [F.relu(conv(x)) for conv in self.convs]

[0361] pooled = [F.max_pool1d(out, out.shape[2]).squeeze(2) for out in conv_outputs]

[0362] cat = torch.cat(pooled, dim=1)

[0363] out = self.fc(cat)

[0364] return out

[0365] ```

[0366] c) TCN Model:

[0367] ```python

[0368] class TCNModel(nn.Module):

[0369] def __init__(self, input_dim, num_channels, kernel_size, output_dim):

[0370] super(TCNModel, self).__init__()

[0371] self.tcn = TemporalConvNet(input_dim, num_channels, kernel_size)

[0372] self.fc = nn.Linear(num_channels[-1], output_dim)

[0373] def forward(self, x):

[0374] x = x.transpose(1, 2) # (batch, input_dim, time_steps)

[0375] out = self.tcn(x)

[0376] out = self.fc(out[:, :, -1])

[0377] return out

[0378] ```

[0379] 4. Train each model

[0380] For each model, the embodiment of the present invention uses the same training data for training. Here is a simplified training process:

[0381] ```Python

[0382] def train_model(model, train_loader, optimizer, criterion, num_epochs):

[0383] for epoch in range(num_epochs):

[0384] for batch_X, batch_y in train_loader:

[0385] optimizer.zero_grad()

[0386] outputs = model(batch_X)

[0387] loss = criterion(outputs, batch_y)

[0388] loss.backward()

[0389] optimizer.step()

[0390] return model

[0391] # Train each model

[0392] lstm_model = train_model(LSTMModel(...), train_loader, optimizer,criterion, num_epochs)

[0393] cnn_model = train_model(CNNModel(...), train_loader, optimizer,criterion, num_epochs)

[0394] tcn_model = train_model(TCNModel(...), train_loader, optimizer,criterion, num_epochs)

[0395] ```

[0396] 5. Ensemble Learning

[0397] The present invention uses a simple average method to integrate the predictions of these models:

[0398] ```Python

[0399] def ensemble_predict(models, X):

[0400] predictions = [model(X) for model in models]

[0401] return torch.mean(torch.stack(predictions), dim=0)

[0402] models = [lstm_model, cnn_model, tcn_model]

[0403] ensemble_prediction = ensemble_predict(models, test_X)

[0404] ```

[0405] 6. Evaluating the Ensemble Model

[0406] The present invention can use the root mean square error (RMSE) to evaluate the model performance:

[0407] ```Python

[0408] from sklearn.metrics import mean_squared_error

[0409] import numpy as np

[0410] def evaluate_model(model, test_X, test_y):

[0411] predictions = model(test_X).detach().numpy()

[0412] return np.sqrt(mean_squared_error(test_y, predictions))

[0413] ensemble_rmse = evaluate_model(lambda x: ensemble_predict(models, x),test_X, test_y)

[0414] print(f"Ensemble Model RMSE: {ensemble_rmse}")

[0415] ```

[0416] 7. Interpret the results

[0417] The embodiments of the present invention obtain the following results:

[0418] ```

[0419] LSTM Model RMSE: 156.3 hours

[0420] CNN Model RMSE: 148.7 hours

[0421] TCN Model RMSE: 152.1 hours

[0422] Ensemble Model RMSE: 143.5 hours

[0423] ```

[0424] This shows that the performance of the ensemble model is better than any single model, demonstrating the effectiveness of deep ensemble learning.

[0425] 8. Applied to maintenance decision of depaneling machine

[0426] Predict maintenance time: Use an integrated model to predict RUL, and trigger a maintenance alert when the predicted value falls below a certain threshold (such as 200 hours).

[0427] Develop maintenance plans: Based on RUL predictions, optimize maintenance plans to avoid unnecessary downtime.

[0428] Fault diagnosis: Analyze the differences in predictions from different models to reveal different types of potential failure modes.

[0429] Performance Monitoring: Continuously monitor the prediction error to assess the overall health of the depaneler.

[0430] Through this deep integrated learning method, the embodiment of the present invention not only improves the accuracy of prediction, but also enhances the robustness of the model. Different types of deep learning models can capture different features and patterns in the data, and their combination provides a more comprehensive and reliable prediction of the remaining service life of the board splitter. This provides a powerful tool for formulating intelligent maintenance strategies, improving equipment utilization and reducing maintenance costs.

[0431] S3.4: Processing the spatiotemporal interference of the multiple deep learning models, identifying and processing spatial and temporal interference factors.

[0432] Spatiotemporal interference comes from multiple sources, such as the spatial distribution of sensors, changes in different time scales, etc. To deal with these interferences, the following strategies can be adopted:

[0433] 1. Spatial interference handling: Use spatial attention mechanism or graph convolutional network to capture the spatial relationship between sensors.

[0434] 2. Temporal interference processing: Use multi-scale temporal convolution or attention mechanism to handle changes at different time scales.

[0435] 3. Spatiotemporal joint modeling: Use spatiotemporal convolutional networks or spatiotemporal attention mechanisms to simultaneously handle interference in spatial and temporal dimensions.

[0436] S3.5: Construct a causal relationship model based on the processed multiple deep learning models.

[0437] In this step, the processed deep learning model output is combined with the previously established causal structure to build a comprehensive causal model. This model can not only capture complex nonlinear relationships, but also maintain causal interpretability. This can be achieved using frameworks such as structural causal models (SCM) or dynamic Bayesian networks (DBN).

[0438] S4: Based on the results of the off-policy learning, extract the features of the sensor data and generate causal sequence data.

[0439] This step is mainly to further process and extract features from the causal sequence data generated in S3. The following methods can be used:

[0440] 1. Time series feature extraction: Use statistical features (such as mean, variance, skewness, kurtosis, etc.) and frequency domain features (such as Fourier transform coefficients) to describe the characteristics of the time series.

[0441] 2. Causal strength features: Based on the estimated results of the causal effect, features reflecting the strength of the causal relationship are extracted.

[0442] 3. Temporal pattern features: Use dynamic time warping (DTW) or sequential pattern mining algorithms to extract temporal pattern features.

[0443] 4. Anomaly detection features: Based on the results of off-policy learning, features that can reflect potential anomalies are extracted.

[0444] Wherein, based on the results of the off-policy learning, causal sequence data is generated, including: identifying key causal chains, building temporal causal relationships according to estimated causal effects, and generating sequence data containing causal relationship strength and temporal information.

[0445] This step transforms the results of off-policy learning into structured causal sequence data. First, the key causal chains are identified based on the estimated causal effects, which represent the causal relationships that have the greatest impact on the performance and failure of the depaneling machine. Then, these causal relationships are organized into a temporal structure that reflects the dynamic relationship between variables over time. Finally, sequence data containing causal strength and temporal information are generated, which will serve as an important input for subsequent modeling.

[0446] S5: Based on the extracted features and the causal sequence data, the hybrid model is applied to monitor the status of the board splitter in real time, predict potential failures, and generate maintenance suggestions.

[0447] The hybrid model includes a hybrid convolutional neural network CNN-Transformer and a state space model. Before the hybrid model monitors the state of the board splitter, the following steps are also included:

[0448] S5.1: Construct a physical model and analyze the physical characteristics of the panel splitter.

[0449] The construction of the physical model requires a deep understanding of the working principle and physical characteristics of the depaneling machine. This includes:

[0450] 1. Dynamic model: describes the motion characteristics of the main components of the panel splitter (such as spindle and cutting device).

[0451] 2. Thermodynamic model: describes the heat generation and transfer during the operation of the equipment.

[0452] 3. Materials science models: describe the deformation and fracture behavior of materials during cutting.

[0453] 4. Friction and wear model: describes the wear process of key components (such as bearings and cutting tools).

[0454] These physical models can be expressed by ordinary differential equations or partial differential equations, providing a theoretical basis for subsequent state-space models.

[0455] S5.2: Extract spatiotemporal features from the sensor data using CNN.

[0456] S5.2.1: Extract local features of the sensor data through multiple convolutional layers.

[0457] Convolutional layers can effectively capture local patterns and features of data. For multi-sensor data, 3D convolution can be used to process both time and space dimensions. For example, a network structure containing multiple 3D convolutional layers can be designed, with each layer using convolution kernels of different sizes to capture features of different scales.

[0458] S5.2.2: Use pooling layers to reduce the dimensionality of the local features.

[0459] Pooling layers (such as max pooling or average pooling) can reduce the dimensionality of feature maps, improving computational efficiency while retaining important information. In 3D convolutional networks, 3D pooling operations can be used.

[0460] S5.2.3: Apply normalization layers to speed up training and improve the stability of the CNN model.

[0461] Batch Normalization or Layer Normalization can be used to speed up the training process and improve the stability of the model. These techniques reduce internal covariate shift by standardizing the input to each layer, making the model easier to train.

[0462] S5.2.4: Integrate feature information through fully connected layers.

[0463] The last one or several fully connected layers are used to integrate the features extracted by the previous convolutional layers and pooling layers to generate higher-level abstract representations. The output of these fully connected layers can be used as the input of the subsequent Transformer model.

[0464] S5.3: Apply the Transformer architecture to perform sequence modeling on the extracted spatiotemporal features to capture long-term dependencies.

[0465] S5.3.1: The spatiotemporal features are converted into a format corresponding to the Transformer through the input embedding layer to form a spatiotemporal feature sequence.

[0466] The input embedding layer converts the features extracted by the CNN into a format suitable for processing by the Transformer. This usually involves mapping multi-dimensional features to a vector space of fixed dimension. For example, if the output of the CNN is a tensor of shape (batch_size, time_steps, feature_dim), it can be mapped to the form of (batch_size, time_steps, d_model) through a linear transformation, where d_model is the model dimension of the Transformer.

[0467] S5.3.2: Apply a multi-head self-attention mechanism to capture the dependencies of the spatiotemporal feature sequences.

[0468] The multi-head self-attention mechanism is a core component of the Transformer, which allows the model to focus on different positions in the sequence and capture complex dependencies. For the spatiotemporal feature sequence of the depaneling machine, multi-head attention can simultaneously learn the relationship between different time scales and different sensors.

[0469] In specific implementation, the attention weight can be calculated using the following formula:

[0470] Attention(Q, K, V) = softmax(QK T / d_k)V

[0471] Where Q, K, V represent query, key and value matrices respectively, and d_k is the dimension of the key vector. The multi-head attention mechanism enhances the expressiveness of the model by computing multiple such attention functions in parallel.

[0472] S5.3.3: Use feedforward neural networks to enhance the nonlinear expression capabilities of Transformer.

[0473] In each Transformer layer, the multi-head self-attention is followed by a feed-forward neural network (FFN). This FFN usually contains two linear transformations and a nonlinear activation function:

[0474] FFN(x) = max(0, xW_1 + b_1)W_2 + b_2

[0475] Where W_1, W_2, b_1 and b_2 are learnable parameters, and max(0, x) represents the ReLU activation function. This FFN enhances the model's ability to capture complex nonlinear relationships.

[0476] S5.3.4: Add position coding to provide position information of the spatiotemporal feature sequence.

[0477] Positional encodings are crucial for the Transformer model to understand the order information of a sequence. Positional encodings can be generated using sine and cosine functions:

[0478] PE(pos, 2i) = sin(pos / 10000^(2i / d_model))

[0479] PE(pos, 2i+1) = cos(pos / 10000^(2i / d_model))

[0480] Where pos is the position and i is the dimension. This encoding allows the model to learn to pay attention to relative positions.

[0481] S5.4: Integrate the state space model and combine the constraints of the physical model to fuse and optimize the hybrid model.

[0482] S5.4.1: Define key state variables, including spindle angular velocity, temperature, vibration amplitude, and tool wear.

[0483] These key state variables directly reflect the working status and health of the board splitter. For example:

[0484] Spindle angular speed: reflects cutting efficiency and load conditions;

[0485] Temperature: Indicates the thermal status of the equipment. Too high a temperature may cause material deformation or equipment damage.

[0486] Vibration amplitude: reflects the stability and wear degree of mechanical parts;

[0487] Tool wear: directly affects cutting quality and efficiency;

[0488] S5.4.2: Establish state transfer equations and observation equations, where the state transfer equations describe the evolution of state variables over time and the observation equations describe the relationship between state variables and observed data.

[0489] The state transfer equation can be expressed as:

[0490] x(t+1) = f(x(t), u(t)) + w(t)

[0491] where x(t) is the state vector at time t, u(t) is the control input, and w(t) is the process noise.

[0492] The observation equation can be expressed as:

[0493] y(t) = h(x(t)) + v(t)

[0494] where y(t) is the observation vector, h(·) is the observation function, and v(t) is the observation noise.

[0495] S5.4.3: Convert the dynamic equations in the physical model into a state-space representation.

[0496] S5.4.3.1: Discretize the continuous-time dynamical equations.

[0497] Discretize the continuous-time equation using a numerical method such as Euler's method or Runge-Kutta's method. For example, using the forward Euler method:

[0498] x(t+Δt) ≈ x(t) + Δt * f(x(t), u(t))

[0499] where Δt is the time step.

[0500] S5.4.3.2: Map physical parameters to parameters of the state-space model.

[0501] Mapping the parameters in the physical model (such as mass, elastic coefficient, friction coefficient, etc.) to the corresponding parameters of the state-space model. This requires a parameter identification or optimization process.

[0502] S5.4.3.3: Design regularization terms based on physical constraints to ensure that predictions are consistent with physical laws.

[0503] Add physical laws-based regularization terms to the model's loss function. For example, you can add energy conservation or momentum conservation constraints:

[0504] L = L_data + λ * L_physics

[0505] Where L_data is the data-driven loss term, L_physics is the regularization term based on physical laws, and λ is the trade-off coefficient.

[0506] S5.4.4: Implement a hybrid filtering algorithm that combines Transformer outputs and physical model predictions to obtain state estimates and predictions.

[0507] S5.4.4.1: Use an extended Kalman filter or an unscented Kalman filter as the filter.

[0508] For nonlinear systems, the Extended Kalman Filter (EKF) or the Unscented Kalman Filter (UKF) are suitable choices. UKF usually performs better when dealing with strongly nonlinear systems.

[0509] S5.4.4.2: The output of the Transformer is input as observations into the filter.

[0510] The output of the Transformer can be viewed as a high-level observation of the system state. This output is input into the filter as y(t) in the observation equation.

[0511] S5.4.4.3: Use physical model predictions as prior information.

[0512] The predictions of the physical model can be used as a prior estimate of the filter. This can be achieved by using the physical model in the prediction step.

[0513] S5.4.4.4: Obtain optimal state estimates and predictions through filtering algorithms.

[0514] The filtering algorithm alternates between prediction and update steps, combining physical model predictions with Transformer outputs to obtain the optimal state estimate.

[0515] Example: Real-time status monitoring and fault prediction system for depaneling machines

[0516] 1. System input

[0517] The embodiment of the present invention has the following real-time data stream:

[0518] Sensor data (collected once every minute):

[0519] Spindle speed (rpm)

[0520] Cutting force (N)

[0521] Vibration amplitude (mm)

[0522] Temperature (°C)

[0523] Current (A)

[0524] Historical maintenance records;

[0525] The causal sequence data generated in the previous step;

[0526] 2. Hybrid Model Architecture

[0527] The hybrid model of the embodiment of the present invention includes:

[0528] a) CNN-Transformer: used to extract spatiotemporal features and long-term dependencies;

[0529] b) State space model: Combined with physical model constraints, state estimation and prediction are performed;

[0530] 3. CNN-Transformer model implementation

[0531] ```Python

[0532] import torch

[0533] import torch.nn as nn

[0534] class CNNTransformer(nn.Module):

[0535] def __init__(self, input_dim, num_filters, num_heads, ff_dim, output_dim):

[0536] super(CNNTransformer, self).__init__()

[0537] self.conv = nn.Conv1d(input_dim, num_filters, kernel_size=3, padding=1)

[0538] self.transformer = nn.TransformerEncoderLayer(d_model=num_filters,nhead=num_heads)

[0539] self.fc = nn.Linear(num_filters, output_dim)

[0540] def forward(self, x):

[0541] # x shape: (batch_size, sequence_length, input_dim)

[0542] x = x.permute(0, 2, 1) # CNN expects (batch_size, input_dim,sequence_length)

[0543] x = self.conv(x)

[0544] x = x.permute(2, 0, 1) # Transformer expects (sequence_length, batch_size, num_filters)

[0545] x = self.transformer(x)

[0546] x = x.mean(dim=0) # Global average pooling

[0547] return self.fc(x)

[0548] model = CNNTransformer(input_dim=5, num_filters=64, num_heads=4, ff_dim=256, output_dim=3)

[0549] ```

[0550] 4. State Space Model

[0551] Use a simplified linear state space model:

[0552] ```python

[0553] import numpy as np

[0554] from filterpy.kalman import KalmanFilter

[0555] class SimplifiedStateSpaceModel:

[0556] def __init__(self, dim_x, dim_z):

[0557] self.kf = KalmanFilter(dim_x=dim_x, dim_z=dim_z)

[0558] # Initialize state transfer matrix, observation matrix, etc.

[0559] self.kf.F = np.eye(dim_x) # state transfer matrix

[0560] self.kf.H = np.eye(dim_z, dim_x) # Observation matrix

[0561] def predict(self):

[0562] self.kf.predict()

[0563] return self.kf.x

[0564] def update(self, z):

[0565] self.kf.update(z)

[0566] return self.kf.x

[0567] ssm = SimplifiedStateSpaceModel(dim_x=3, dim_z=3)

[0568] ```

[0569] 5. Real-time monitoring and prediction process

[0570] ```Python

[0571] def monitor_and_predict(sensor_data, cnn_transformer, ssm):

[0572] # Extract features using CNN-Transformer

[0573] features = cnn_transformer(sensor_data)

[0574] # State space model prediction

[0575] predicted_state = ssm.predict()

[0576] # Update state estimate

[0577] updated_state = ssm.update(features.detach().numpy())

[0578] return updated_state

[0579] # This embodiment of the present invention has a function to obtain the latest sensor data

[0580] def get_latest_sensor_data():

[0581] # This should be to get data from the actual sensor

[0582] return torch.randn(1, 60, 5) # This embodiment of the present invention has 5-dimensional sensor data with 60 time steps

[0583] # Real-time monitoring loop

[0584] while True:

[0585] sensor_data = get_latest_sensor_data()

[0586] current_state = monitor_and_predict(sensor_data, model, ssm)

[0587] # Analyze the current status

[0588] analyze_state(current_state)

[0589] # Wait for the next data point

[0590] time.sleep(60) # execute once every minute

[0591] ```

[0592] 6. Status analysis and fault prediction

[0593] ```Python

[0594] def analyze_state(state):

[0595] # The state vector contains: [spindle health, cutting system health, overall system health]

[0596] main_shaft_health, cutting_system_health, overall_health = state

[0597] # Define health threshold

[0598] HEALTH_THRESHOLD = 0.7

[0599] WARNING_THRESHOLD = 0.8

[0600] if overall_health < HEALTH_THRESHOLD:

[0601] print("Warning: The overall health of the system is low, a comprehensive check is recommended")

[0602] generate_maintenance_advice("full_inspection")

[0603] elif main_shaft_health < WARNING_THRESHOLD:

[0604] print("Note: The spindle health has deteriorated and maintenance is required")

[0605] generate_maintenance_advice("main_shaft_maintenance")

[0606] elif cutting_system_health < WARNING_THRESHOLD:

[0607] print("Note: The health of the cutting system has decreased. It is recommended to check the cutting tool")

[0608] generate_maintenance_advice("cutting_tool_inspection")

[0609] else:

[0610] print("The system is running normally")

[0611] def generate_maintenance_advice(issue):

[0612] advice_map = {

[0613] "full_inspection": "Please arrange for a full inspection to check all critical components. Recommended time for completion is 48 hours.",

[0614] "main_shaft_maintenance": "It is recommended to inspect and lubricate the main shaft bearings at the next scheduled downtime. This is expected to take 4 hours.",

[0615] "cutting_tool_inspection": "Please replace the cutting tool after the current batch is completed. Estimated time: 1 hour."

[0616] }

[0617] print("Maintenance suggestion:", advice_map.get(issue, "Please contact the maintenance team for further diagnosis"))

[0618] ```

[0619] 7. Example Run

[0620] After the system is run for a period of time, the following output is obtained:

[0621] ```

[0622] The system is operating normally;

[0623] The system is operating normally;

[0624] Note: The health of the cutting system has decreased, it is recommended to check the cutting tool;

[0625] Maintenance advice: Please replace the cutting tool after the current batch is completed. Estimated time: 1 hour.

[0626] The system is operating normally;

[0627] Warning: The overall health of the system is low, and a comprehensive check is recommended;

[0628] Maintenance Recommendation: Please arrange for a full overhaul to check all critical components. This is recommended to be completed within 48 hours.

[0629] ```

[0630] 8. Interpretation and Application of Results

[0631] Real-time monitoring: The system can monitor the health status of each key component of the depaneling machine in real time.

[0632] Early warning: Before problems occur in the cutting system, the system gives early warning and specific maintenance suggestions.

[0633] Preventive maintenance: By predicting potential failures, the system is able to recommend appropriate maintenance actions, such as replacing cutting tools.

[0634] Comprehensive diagnostics: When a decline in overall health is detected, the system recommends a comprehensive overhaul to help identify and resolve potential systemic issues.

[0635] Maintenance plan optimization: Based on the generated maintenance recommendations, the maintenance team can better schedule maintenance time to minimize the impact on production.

[0636] 9. Further Improvements

[0637] Dynamic threshold: Dynamically adjust the health threshold based on historical data and machine learning models.

[0638] Multi-sensor fusion: Integrate more types of sensor data to improve the accuracy of state estimation.

[0639] Prediction time domain expansion: not only predict the current state, but also predict the state change trend in the future.

[0640] Maintenance effect feedback: Record the system performance improvement after each maintenance operation to optimize future maintenance decisions.

[0641] Enhanced explainability: Provides more detailed status analysis reports to help maintainers understand the decision-making basis of the system.

[0642] Through this example, the embodiment of the present invention shows how to combine the deep learning model (CNN-Transformer) with the state space model to achieve real-time state monitoring, fault prediction and maintenance suggestion generation of the depaneling machine. This hybrid method not only takes advantage of the advantages of deep learning in processing complex spatiotemporal data, but also combines the interpretability of the state space model and the consideration of physical constraints, providing a powerful and flexible solution for the intelligent maintenance of the depaneling machine.

[0643] Through the above steps, this method realizes real-time monitoring of the status of the depaneling machine, prediction of potential faults, and generation of maintenance suggestions. This method combines the advantages of data-driven methods and physical models, which can capture complex nonlinear relationships while maintaining physical interpretability, thereby improving the accuracy and reliability of fault prediction and diagnosis.

[0644] The method and system of this application achieves comprehensive monitoring and accurate prediction of the operating status of the board splitter by fusing multi-sensor data, using causal inference and deep learning technology, and combining physical model constraints. This not only improves the reliability and production efficiency of the equipment, but also provides new ideas and solutions for the intelligent upgrading of the manufacturing industry. Embodiment 2

[0645] like Figure 2 As shown, the embodiment of the present application also provides an intelligent multi-sensor fusion fault prediction and diagnosis system 20 for a board splitter, comprising:

[0646] The acquisition module 21 is used to collect data from various sensors on the board separation machine;

[0647] A preprocessing module 22, used for performing random smoothing secondary sampling preprocessing on the sensor data;

[0648] An off-policy learning module 23, used to establish a causal relationship model using a causal depth set of off-policy evaluation under spatiotemporal interference, and perform off-policy learning of the sensor data;

[0649] A generating module 24, configured to extract features of the sensor data and generate causal sequence data based on the results of the off-policy learning;

[0650] The monitoring module 25 is used to monitor the state of the board splitter in real time based on the extracted features and the causal sequence data, and to predict potential failures and generate maintenance suggestions by applying a hybrid model. Embodiment 3

[0651] The present application also provides a computer device, such as Figure 3 As shown, the computer device comprises:

[0652] at least one processor; and,

[0653] a memory communicatively connected to the at least one processor; wherein,

[0654] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned intelligent multi-sensor fusion board splitter fault prediction and diagnosis method.

[0655] An embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the above-mentioned intelligent multi-sensor fusion board splitter fault prediction and diagnosis method.

[0656] An embodiment of the present application also provides a computer program product, including computer instructions, characterized in that when the computer instructions are executed by a processor, the steps of the above-mentioned intelligent multi-sensor fusion board splitter fault prediction and diagnosis method are implemented.

[0657] Each module in the above intelligent multi-sensor fusion fault prediction and diagnosis system for the board splitter can be implemented in whole or in part by software, hardware and their combination. Each module can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the corresponding operations of each module.

[0658] In one embodiment, a computer device is also provided, which is the intelligent multi-sensor fusion fault prediction and diagnosis system for the board splitter mentioned in the above method embodiment. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface.

[0659] Among them, the processor of the computer device is used to provide computing and control capabilities, and can be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, etc., but is not limited thereto. The processor may include one or more processors, for example, one or more central processing units (CPUs). When the processor is a CPU, the CPU may be a single-core CPU or a multi-core CPU. The processor may also include one or more dedicated processors, and the dedicated processor may include a GPU, an FPGA, etc., for accelerated processing. The processor is used to call the program code and data in the memory and execute the steps in the above method embodiment. For details, please refer to the description in the method embodiment, which will not be repeated here.

[0660] The memory of the computer device includes, but is not limited to, non-volatile storage media and internal memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium.

[0661] The input / output interface of the computer device is used to exchange information between the processor and external devices.

[0662] The communication interface of the computer device is used to communicate with an external terminal via a network connection.

[0663] When the computer program is executed by a processor, the above method is implemented.

[0664] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the division of each unit / module is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The mutual coupling, direct coupling, or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of systems or units, which can be electrical, mechanical or other forms.

[0665] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0666] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a read-only memory (ROM), or a random access memory (RAM), or a magnetic medium, such as a floppy disk, a hard disk, a tape, a disk, or an optical medium, such as a digital versatile disc (DVD), or a semiconductor medium, such as a solid state disk (SSD), etc.

[0667] The system of the embodiments of the present disclosure can execute the method provided by the embodiments of the present disclosure, and the implementation principles are similar. The actions performed by each module in the system of each embodiment of the present disclosure correspond to the steps in the method of each embodiment of the present disclosure. For the detailed functional description of each module of the system, please refer to the description in the corresponding method shown in the previous text, which will not be repeated here.

[0668] The above is only an optional implementation method for some implementation scenarios of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present disclosure, other similar implementation methods based on the technical ideas of the present disclosure are also within the protection scope of the embodiments of the present disclosure.

[0669] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present application, and these modifications or replacements should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. An intelligent multi-sensor fusion fault prediction and diagnosis method for a board splitter, characterized in that: The following steps are involved: Collect data from various sensors on the depaneling machine; Perform random smoothing and subsampling preprocessing on sensor data; Establishing a causal relationship model using a causal depth set of off-policy evaluations under spatiotemporal interference, and performing off-policy learning of the sensor data; Based on the results of the off-policy learning, extracting features of the sensor data and generating causal sequence data; Based on the extracted features and the causal sequence data, a hybrid model is applied to monitor the status of the depaneling machine in real time, predict potential failures, and generate maintenance suggestions; The method of establishing a causal relationship model using the causal depth set of off-strategy evaluation under spatiotemporal interference includes: Identify key variables and establish causal relationships between variables; Perform off-policy learning to learn the optimal strategy from historical data; Perform deep ensemble learning to train and integrate multiple deep learning models based on the optimal strategy; Processing the spatial and temporal interference of the multiple deep learning models, identifying and processing spatial and temporal interference factors; Building a causal relationship model based on the processed multiple deep learning models; Among them, off-policy learning includes: Collecting historical data, wherein the historical data includes historical sensor data, maintenance records, and fault logs; Define the processing process and processing results, where the processing process includes maintenance operations and parameter adjustments, and the processing results include equipment performance indicators and fault status; Estimation of propensity scores, including: using logistic regression or random forests to calculate the probability of taking either treatment given covariates; Construct counterfactual regression models to predict outcomes under different treatments; The weighted mean treatment effect was calculated using the inverse probability weighting method; The double robust method is used to combine regression model and inverse probability weighting to estimate causal effects, which include calculating average treatment effect and conditional average treatment effect. Confidence intervals for causal effects were estimated using bootstrapping; Based on the results of the off-policy learning, causal sequence data is generated, including: identifying key causal chains, building temporal causal relationships based on estimated causal effects, and generating sequence data containing causal relationship strength and temporal information.

2. The method according to claim 1, characterized in that The random smoothing secondary sampling preprocessing of the sensor data includes: Performing initial sampling on the sensor data; Perform random smoothing on the initial sampling data; The smoothed initial sampling data is sub-sampled.

3. The method according to claim 2, characterized in that The random smoothing process of the initial sampling data comprises: Define the smoothing function; In the initial sampled data, generating a random disturbance for each data point; The smoothing function is applied to process the randomly disturbed data.

4. The method according to claim 1, characterized in that: The hybrid model includes a hybrid convolutional neural network CNN-Transformer and a state space model. Before the hybrid model is applied to monitor the state of the board splitter in real time, the method further includes: Build a physical model to analyze the physical characteristics of the panel splitter; Extracting spatiotemporal features from the sensor data using CNN; Applying the Transformer architecture to perform sequence modeling on the extracted spatiotemporal features to capture long-term dependencies; The state space model is integrated and combined with the constraints of the physical model to fuse and optimize the hybrid model.

5. The method according to claim 4, characterized in that CNN is used to extract spatiotemporal features from the sensor data, including: Extracting local features of the sensor data through multiple convolutional layers; Using a pooling layer to reduce the dimension of the local features; Applying normalization layers to speed up training and improve the stability of the CNN; The feature information is integrated through the fully connected layer.

6. The method according to claim 4, characterized in that The Transformer architecture is used to perform sequence modeling on the extracted spatiotemporal features, including: The spatiotemporal features are converted into a format corresponding to the Transformer through an input embedding layer to form a spatiotemporal feature sequence; Applying a multi-head self-attention mechanism to capture the dependencies of the spatiotemporal feature sequences; Use feedforward neural networks to enhance the nonlinear expression capabilities of Transformer; Position coding is added to provide position information of the spatiotemporal feature sequence.

7. The method according to claim 4, characterized in that The integrated state space model, combined with the constraints of the physical model, includes: Define key state variables, including spindle angular velocity, temperature, vibration amplitude, and tool wear; Establish state transfer equations and observation equations, where the state transfer equations describe the evolution of state variables over time, and the observation equations describe the relationship between state variables and observed data; The dynamic equations in the physical model are converted into state space representation, including: Discretize the continuous-time dynamic equations; Mapping physical parameters to parameters of the state-space model; Design regularization terms based on physical constraints to ensure that predictions conform to physical laws; Implement a hybrid filtering algorithm that combines Transformer output and physical model prediction to obtain state estimation and prediction, including: using an extended Kalman filter or an unscented Kalman filter as a filter; inputting the Transformer output as an observation value into the filter; using the physical model prediction as prior information; and obtaining the optimal state estimation and prediction through a filtering algorithm.

8. An intelligent multi-sensor fusion fault prediction and diagnosis system for a board splitter, characterized in that: include: The acquisition module is used to collect data from various sensors on the depaneling machine; A preprocessing module is used to perform random smoothing and secondary sampling preprocessing on sensor data; An off-policy learning module, for establishing a causal relationship model using a causal depth set of off-policy evaluation under spatiotemporal interference, and performing off-policy learning of the sensor data; A generation module, configured to extract features of the sensor data and generate causal sequence data based on the results of the off-policy learning; A monitoring module, for monitoring the state of the board splitter in real time by applying a hybrid model based on the extracted features and the causal sequence data, and predicting potential failures and generating maintenance suggestions; The method of establishing a causal relationship model using the causal depth set of off-strategy evaluation under spatiotemporal interference includes: Identify key variables and establish causal relationships between variables; Perform off-policy learning to learn the optimal strategy from historical data; Perform deep ensemble learning to train and integrate multiple deep learning models based on the optimal strategy; Processing the spatial and temporal interference of the multiple deep learning models, identifying and processing spatial and temporal interference factors; Building a causal relationship model based on the processed multiple deep learning models; Among them, off-policy learning includes: Collecting historical data, wherein the historical data includes historical sensor data, maintenance records, and fault logs; Define the processing process and processing results, where the processing process includes maintenance operations and parameter adjustments, and the processing results include equipment performance indicators and fault status; Estimation of propensity scores, including: using logistic regression or random forests to calculate the probability of taking either treatment given covariates; Construct counterfactual regression models to predict outcomes under different treatments; The weighted mean treatment effect was calculated using the inverse probability weighting method; The double robust method is used to combine regression model and inverse probability weighting to estimate causal effects, which include calculating average treatment effect and conditional average treatment effect. Confidence intervals for causal effects were estimated using bootstrapping; Based on the results of the off-policy learning, causal sequence data is generated, including: identifying key causal chains, building temporal causal relationships based on estimated causal effects, and generating sequence data containing causal relationship strength and temporal information.

Images