A time sequence feature decomposition and dynamic fusion workshop man-hour prediction method, device, medium and system

Abstract

This invention discloses a method, device, medium, and system for predicting workshop working hours based on temporal feature decomposition and dynamic fusion, belonging to the field of workshop management technology. The method includes the following steps: multi-source data acquisition and preprocessing; receiving multi-source heterogeneous data related to workshop field working hours; performing three-component decomposition of the preprocessed data into a working hour time series; predicting the decomposed data using the linear component of the ARIMA model and the nonlinear perturbation prediction using the LSTM model; and performing adaptive weighted fusion based on real-time fluctuation intensity on the predicted data. This invention achieves accurate extraction and dynamic fusion of linear temporal features and nonlinear deep features, improving the accuracy and applicability of field working hour prediction and enhancing the accuracy of working hour prediction.

Description

Technical Field

[0001] This invention relates to the field of workshop management technology, and more specifically, to a method, device, medium, and system for predicting workshop working hours by decomposing and dynamically fusing time-series features. Background Technology

[0002] Monthly working hours in the workshop field are core data for enterprise production scheduling, staffing, and cost accounting. Their time series exhibits both linear trends and periodic fluctuations, as well as nonlinear random disturbances. Currently, two main technical solutions are used for predicting workshop field working hours: (1) Traditional statistical forecasting methods: Represented by the ARIMA (Autoregressive Integral Moving Average) model, this model is a classic linear time series forecasting method. Through autoregression, differencing, and moving average operations, it can effectively capture the linear trend and periodic pattern in the time series. The model has a simple structure, strong interpretability, and high computational efficiency, and is suitable for forecasting time series data with strong stability. The shortcomings of the single ARIMA model: The working hours in the workshop are affected by multiple nonlinear and sudden factors such as changes in production orders. The time series data has strong perturbation. ARIMA can only fit the linear pattern and cannot capture the coupling effect of nonlinear random fluctuations and external factors. The prediction error is large and the robustness is poor, making it difficult to meet the needs of fine chemical time management in the workshop.

[0003] (2) Deep learning prediction methods: Represented by LSTM (Long Short-Term Memory Network), as an improved recurrent neural network, it can mine long-distance dependencies in time series data through the gating mechanism of forget gate, input gate, and output gate. It has excellent nonlinear feature fitting ability and can handle time series data with complex perturbations and multi-factor coupling. The shortcomings of a single LSTM model are: low fitting efficiency for explicit linear trends and periodic patterns in work time series, model training depends on a large number of samples, and it is prone to overfitting in small sample scenarios. Moreover, the model has weak interpretability and cannot provide clear regularity for workshop production scheduling and work time management.

[0004] Some existing technical solutions attempt to simply combine ARIMA and LSTM, generally using a residual superposition method: first, ARIMA is used to predict the original work time series, then the residuals generated by the ARIMA prediction are input into the LSTM model for fitting, and finally the two results are added together to obtain the final predicted value. The shortcomings of traditional simple residual fusion models are: they only use LSTM as a residual correction tool for ARIMA, without performing time-series decomposition for the characteristics of workshop field work time data; the feature extraction of the two models interferes with each other; and the fixed weight superposition method cannot adaptively adjust the model weights according to the real-time fluctuations of the work time series, resulting in limited improvement in prediction accuracy and an inability to adapt to scenarios with dynamic changes in workshop field work times. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a workshop time prediction method, device, medium and system for time series feature decomposition and dynamic fusion, so as to realize the accurate extraction and dynamic fusion of linear time series features and nonlinear deep features, improve the accuracy and applicability of field time prediction and improve the accuracy of time prediction.

[0006] The objective of this invention is achieved through the following solution: A method for predicting shop floor time using temporal feature decomposition and dynamic fusion includes the following steps: Multi-source data acquisition and preprocessing, receiving multi-source heterogeneous data related to workshop and field working hours; The preprocessed data is then decomposed into three components of the time series of work hours. For the decomposed data, linear component prediction using the ARIMA model and nonlinear perturbation prediction using the LSTM model are employed. For the predicted data, an adaptive weighted fusion based on the real-time fluctuation intensity is performed.

[0007] Furthermore, the multi-source heterogeneous data specifically includes: historical working hour data, field task data, and personnel skill data.

[0008] Furthermore, the preprocessing includes missing value handling, outlier detection, and duplicate value detection. After data cleaning, numerical fields are normalized through data standardization to map the data to the [0, 1] interval. Feature engineering is constructed based on the preprocessed data, where the predictive features include time features, lag features, sliding window features, task features, and personnel features. Through feature engineering, an input matrix X (n×P) containing P features is finally obtained, where n is the number of samples and P is the number of features. At the same time, a target variable Y (n×1) is constructed, which is the working hour value for the corresponding date.

[0009] Furthermore, the three-component decomposition of the work time series specifically includes the following sub-steps: The STL algorithm is used to decompose the preprocessed original field work time series Y(t) into a trend term T(t), a periodic term S(t), and a random fluctuation term R(t). The decomposition formula is as follows: Y(t) = T(t) + S(t) + R(t); Where T(t) represents the long-term trend term, reflecting the long-term linear variation law of the workshop field working hours; S(t) represents the periodic term, reflecting the periodic field law; and R(t) represents the random fluctuation term, reflecting the nonlinear random fluctuation. Adding the trend term and the periodic term together, we get the linear stationary component L(t) = T(t) + S(t), and the random fluctuation term is the nonlinear component N(t) = R(t).

[0010] Furthermore, the linear component prediction of the ARIMA model specifically includes the following sub-steps: The stationarity of the linear stationary component L(t) is tested. The stationarity of the sequence is verified by the ADF method. If the sequence is non-stationary, it is differencing until it becomes stationary. The difference order d is then determined. Next, plot the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots of the stationary series. Combine the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) to determine the autoregression order p and the moving average order q, and construct the ARIMA(p, d, q) model. Then, the ARIMA model is trained and its parameters are optimized using the training dataset to obtain the trained ARIMA model. Finally, the test dataset is input into the trained ARIMA model, which outputs the prediction results of the linear components.

[0011] Furthermore, the nonlinear perturbation prediction of the LSTM model specifically includes the following sub-steps: Constructing a multidimensional nonlinear feature matrix: The nonlinear component N(t) is horizontally concatenated with the collected multidimensional production impact features to obtain a multidimensional input matrix containing temporal disturbance features and external impact features; Construct an LSTM network model with the following structure: input layer, first LSTM layer, dropout layer, second LSTM layer, dropout layer, fully connected layer, and output layer. The model training parameters are set as follows: batch size is 32, number of iterations is 100, the optimizer is Adam, and the loss function is mean squared error (MSE). Next, the LSTM model is trained using the training dataset, and overfitting is prevented by early stopping, resulting in a trained LSTM model. Finally, the test dataset is input into the trained LSTM model, which outputs the prediction results of the nonlinear perturbation.

[0012] Furthermore, the adaptive weighted fusion based on real-time fluctuation intensity specifically includes the following sub-steps: Calculate the real-time fluctuation intensity of the work time series The sliding window method is used to calculate the current time step. The standard deviation of the working hours at each moment is used as the fluctuation intensity, and the formula is: ; in, This represents the average working hours within the sliding window. To adjust the sliding window size; Then establish the mapping relationship between fluctuation intensity and model weights: ; in, These are the output weights of the ARIMA model. The output weights of the LSTM model. and These represent the maximum and minimum values ​​of fluctuation intensity in the training dataset, respectively. Finally, the prediction results of ARIMA and LSTM are adaptively weighted and fused to obtain the final prediction result of field working hours: .

[0013] A workshop time prediction device based on temporal feature decomposition and dynamic fusion includes a processor and a memory. The memory stores a computer program, which, when loaded by the processor, executes the application package multi-level cache distribution method based on topology perception as described above.

[0014] A computer-readable storage medium storing a computer program that, when loaded by a processor, executes the workshop time prediction method of temporal feature decomposition and dynamic fusion as described in any of the preceding claims.

[0015] A workshop time prediction system based on temporal feature decomposition and dynamic fusion includes a workshop time prediction device based on temporal feature decomposition and dynamic fusion as described above.

[0016] The beneficial effects of this invention include: This invention can significantly improve the accuracy of work hour forecasting. Specifically, by capturing time series features and enhancing model interpretability, the ARIMA model can effectively capture the periodicity, trend, and seasonality of linear time series features of work hour data; the LSTM model can capture complex nonlinear features and long-term dependencies; the hybrid model proposed in this invention combines the advantages of both, not only providing high prediction accuracy but also offering model interpretability analysis, supporting medium- and long-term forecasting, assisting resource planning, and supporting monthly and quarterly medium- and long-term work hour forecasting, thus providing decision support for enterprise production planning and human resource scheduling. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an overall flowchart of the method in an embodiment of the present invention; Figure 2This is a comparison chart of the predicted value and the actual value in the method of this embodiment of the invention. Detailed Implementation

[0019] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.

[0020] Field tasks are an important part of workshop production. Reasonable estimation of field task hours allows for the coordinated and efficient allocation of human resources, ensuring the smooth operation of overall workshop production. For example... Figure 1 As shown, this invention specifically relates to improvements in workshop field monthly working hour prediction and time series analysis techniques. In particular, it addresses the technical problems of existing workshop field monthly working hour prediction methods, such as the inability to simultaneously handle linear regularity fitting and nonlinear disturbance capture, rigid model fusion methods, low prediction accuracy, and poor robustness. It provides a workshop field working hour prediction scheme based on time series feature decomposition and dynamic fusion, aiming to achieve accurate extraction and dynamic fusion of linear time series features and nonlinear deep features, thereby improving the accuracy and applicability of field working hour prediction. Specifically, workshop field working hour prediction is performed through time series analysis. In a preferred embodiment, a workshop working hour prediction method based on time series feature decomposition and dynamic fusion is provided, including the following steps: Step 1: Multi-source data acquisition and preprocessing, receiving multi-source heterogeneous data related to workshop and field working hours, including historical working hour data, field task data, and personnel skill data.

[0021] The historical working hours data includes fields such as field time, field task type, actual working hours, and completion status, with a time span of no less than 36 months.

[0022] The field task data includes fields such as task ID, task type (repair / debugging / service / training, etc.), task description, task complexity level, task urgency, target equipment / production line, and customer information.

[0023] The personnel skills data includes fields such as employee ID, employee name, skill level (beginner / intermediate / advanced / expert), professional skill tags, historical field records, and historical working hours performance.

[0024] The collected multi-source data is preprocessed, including data cleaning, such as missing value handling, outlier detection, and duplicate value detection.

[0025] Among the specific implementation methods for handling missing values, for numerical fields, the K-nearest neighbor (KNN) interpolation method is used to fill missing values, and the interpolation error is controlled within 3%; for categorical fields, the mode is used to fill or the field is marked as "unknown" category.

[0026] Among the specific implementation methods of outlier detection, the 3σ criterion is used to detect outliers. The detected outliers are manually reviewed, and after confirmation of anomalies, they are removed or corrected.

[0027] Among the specific implementation methods for handling duplicate values, completely duplicate records are removed, and records with duplicate key fields are merged.

[0028] After data cleaning, data standardization is performed to normalize numerical fields, mapping the data to the [0, 1] interval. The specific normalization formula is as follows: x_normalized=(x-x_min) / (x_max-x_min) Feature engineering is constructed based on preprocessed data. The predictive features include time features, lag features, sliding window features, task features, and personnel features.

[0029] Among them, time characteristics refer to year, month, weekday, weekday / weekend, holiday markers, etc.

[0030] Among them, the lag characteristics refer to the average, maximum, and minimum working hours over the past 1 day, 3 days, 7 days, and 30 days.

[0031] Among them, the sliding window characteristics refer to the average working hours, standard deviation, and rate of change of the 7-day and 30-day sliding windows.

[0032] Among them, task characteristics refer to one-hot encoding of task type, task complexity level, task urgency, etc.

[0033] Personnel characteristics include skill level, historical work hours and performance, and field experience.

[0034] Through feature engineering, an input matrix X (n×P) containing P features is obtained, where n is the number of samples (days) and P is the number of features. Simultaneously, a target variable Y (n×1) is constructed, representing the working hours for the corresponding date.

[0035] Step 2: Three-component decomposition of work time series The preprocessed original field work time series Y(t) was decomposed using the STL (Loess-based seasonal and trend decomposition) algorithm, separating it into a trend term T(t), a periodic term S(t), and a random fluctuation term R(t). The decomposition formula is as follows: Y(t) = T(t) + S(t) + R(t) in: T(t): Long-term trend term, reflecting the long-term linear change of workshop field working hours with enterprise development and production scale.

[0036] S(t): Periodic term, reflecting the periodic external field pattern affected by factors such as production cycle, holidays, and monthly budget.

[0037] R(t): The random fluctuation term reflects the nonlinear random fluctuation caused by unforeseen factors such as sudden equipment failure, urgent customer demand, and temporary project changes.

[0038] Adding the trend term and the periodic term yields the linear stationary component L(t) = T(t) + S(t), while the random fluctuation term serves as the nonlinear component N(t) = R(t).

[0039] Step 3: Prediction of linear components of the ARIMA model The stationarity of the linear stationary component L(t) is tested by using the ADF (unit root test) method to verify the stationarity of the sequence. If the sequence is non-stationary, it is differencing until it becomes stationary, and the difference order d is determined.

[0040] Plot the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots of stationary series, and determine the autoregression order p and the moving average order q by combining AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion), and construct the ARIMA(p, d, q) model.

[0041] The ARIMA model is trained and its parameters are optimized using the training dataset to obtain the trained ARIMA model.

[0042] Input the test dataset into the trained ARIMA model, and output the prediction results of the linear components \hat{L}(t).

[0043] Step 4: Nonlinear perturbation prediction using the LSTM model Construct a multidimensional nonlinear feature matrix: The nonlinear component N(t) is horizontally concatenated with all the multidimensional production impact features collected in step 1 to obtain a multidimensional input matrix containing time-series disturbance features and external impact features.

[0044] Construct an LSTM network model with the following structure: input layer, first LSTM layer, dropout layer, second LSTM layer, dropout layer, fully connected layer, and output layer.

[0045] The model training parameters are set as follows: batch size is 32, number of iterations is 100, the optimizer is Adam, and the loss function is mean squared error (MSE).

[0046] The LSTM model is trained using the training dataset, and overfitting is prevented by early stopping, resulting in a fully trained LSTM model.

[0047] Input the test dataset into the trained LSTM model, and output the prediction result of the nonlinear perturbation t.

[0048] Step 5: Adaptive weighted fusion based on real-time fluctuation intensity Sub-step 5.1: Calculate the real-time fluctuation intensity of the work time series. The sliding window method is used to calculate the current time step. The standard deviation of the working hours at each moment is used as the fluctuation intensity, and the formula is:

[0049] in, This represents the average working hours within the sliding window. This is the size of the sliding window; the default value is 7. Sub-step 5.2: Establish the mapping relationship between fluctuation intensity and model weights:

[0050] in, These are the output weights of the ARIMA model. The output weights of the LSTM model. and These represent the maximum and minimum values ​​of fluctuation intensity in the training dataset, respectively. Sub-step 5.3 involves adaptively weighting and fusing the prediction results of ARIMA and LSTM to obtain the final field work time prediction result: .

[0051] In other embodiments, based on the above embodiments, another method for predicting workshop hours by temporal feature decomposition and dynamic fusion is provided. This method is applied to the prediction of field tasks for a production line, which has an average monthly field work time of over 6000 hours and a field man-day of over 750 man-days. The prediction method for field work time in this embodiment is adopted, and the specific steps are as follows: Step 1: Data Collection Collect historical work hour data from the past 12 months, including: Field date, maintenance task type (preventive maintenance / fault repair), equipment model, equipment age, maintenance personnel skill level, and actual maintenance man-hours; Equipment operating parameters (temperature, vibration spectrum), equipment fault history records; Personnel's historical field records and historical work hours performance.

[0052] Step 2: Data Preprocessing and Feature Engineering Data cleaning: KNN interpolation is used to fill missing values, and the 3σ criterion is used to detect outliers.

[0053] Normalize the data: map all numeric fields to the range [0, 1].

[0054] Features include: time features (month, weekday), lag features (average working hours over the past 1, 3, 7, and 30 days), sliding window features (mean and standard deviation of working hours for 7-day and 30-day sliding windows), task features (task type and task complexity), equipment features (equipment age and equipment failure frequency), personnel features (skill level and historical working performance), and environmental features (temperature and travel distance).

[0055] Step 3: Timing Decomposition The STL algorithm was used to decompose the original work time sequence, with a period length of 30 (monthly period), to obtain a trend term, a periodic term, and a random fluctuation term. The trend term and the periodic term were added together to obtain a linear stationary component, and the random fluctuation term was used as a nonlinear component.

[0056] Step 4: ARIMA Model Training and Prediction An ADF test was performed on the linear stationary component, and the results showed that the sequence was stationary, therefore the difference order (d=0) was determined. Using the ACF and PACF plots combined with the AIC criterion, the autoregressive order (p=2) and the moving average order (q=1) were determined, and an ARIMA(2, 0, 1) model was constructed. After training the model on the training set, predictions were made on the test set to obtain the prediction results for the linear component.

[0057] Step 5: LSTM Model Training and Prediction The nonlinear components are concatenated with five production impact features to form a 6-dimensional input matrix. An LSTM model is constructed: the input layer dimension is 6, the first LSTM layer has 64 neurons with a Dropout rate of 0.2, the second LSTM layer has 32 neurons with a Dropout rate of 0.2, the fully connected layer has 16 neurons, and the output layer dimension is 1. The batch size is set to 32, the number of iterations is 100, and the Adam optimizer is used for training. After training, predictions are made on the test set to obtain the nonlinear perturbation prediction results. Step 5: Anomaly detection and early warning.

[0058] Step 6: Adaptive Weighted Fusion Set the sliding window size (k=7) and calculate the real-time fluctuation intensity of the test set at each time step. Calculate the dynamic weights of ARIMA and LSTM based on the fluctuation intensity, and then perform a weighted fusion of the prediction results from the two models.

[0059] Step 7: Result Verification The fused prediction results are denormalized to obtain the final daily field work hour prediction. Calculations show that the MAE of the method in this invention is 1.23 hours, the RMSE is 1.87 hours, and the MAPE is 8.5%. In comparison, the RMSE of a single ARIMA model is 3.22 hours, the RMSE of a single LSTM model is 2.59 hours, and the RMSE of a traditional residual fusion model is 2.37 hours. The prediction accuracy of the method in this invention is significantly better than existing technologies.

[0060] like Figure 2 As shown, compared with traditional empirical estimation methods, the prediction accuracy of the method in this embodiment of the invention is improved from 65%-70% to over 90%, and the prediction error is reduced from ±15%-±20% to within ±5%. In a 6-month field test data verification, the MAPE (mean absolute percentage error) was controlled within 4.2%.

[0061] It should be noted that, within the scope of protection defined in the claims of this invention, the following embodiments can be combined and / or extended or replaced in any logical manner from the above specific embodiments, such as the disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0062] Example 1 A method for predicting shop floor time using temporal feature decomposition and dynamic fusion includes the following steps: Multi-source data acquisition and preprocessing, receiving multi-source heterogeneous data related to workshop and field working hours; The preprocessed data is then decomposed into three components of the time series of work hours. For the decomposed data, linear component prediction using the ARIMA model and nonlinear perturbation prediction using the LSTM model are employed. For the predicted data, an adaptive weighted fusion based on the real-time fluctuation intensity is performed.

[0063] Example 2 Based on Example 1, the multi-source heterogeneous data specifically includes: historical working hour data, field task data, and personnel skill data.

[0064] Example 3 Based on Example 1, the preprocessing includes missing value handling, outlier detection, and duplicate value detection. After data cleaning, numerical fields are normalized through data standardization to map the data to the [0, 1] interval. Feature engineering is constructed based on the preprocessed data, where the prediction features include time features, lag features, sliding window features, task features, and personnel features. Through feature engineering, an input matrix X (n×P) containing P features is finally obtained, where n is the number of samples and P is the number of features. At the same time, a target variable Y (n×1) is constructed, which is the working hour value for the corresponding date.

[0065] Example 4 Based on Example 3, the three-component decomposition of the working time series specifically includes the following sub-steps: The STL algorithm is used to decompose the preprocessed original field work time series Y(t) into a trend term T(t), a periodic term S(t), and a random fluctuation term R(t). The decomposition formula is as follows: Y(t) = T(t) + S(t) + R(t); Where T(t) represents the long-term trend term, reflecting the long-term linear variation law of the workshop field working hours; S(t) represents the periodic term, reflecting the periodic field law; and R(t) represents the random fluctuation term, reflecting the nonlinear random fluctuation. Adding the trend term and the periodic term together, we get the linear stationary component L(t) = T(t) + S(t), and the random fluctuation term is the nonlinear component N(t) = R(t).

[0066] Example 5 Based on Example 4, the linear component prediction of the ARIMA model specifically includes the following sub-steps: The stationarity of the linear stationary component L(t) is tested. The stationarity of the sequence is verified by the ADF method. If the sequence is non-stationary, it is differencing until it becomes stationary. The difference order d is then determined. Next, plot the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots of the stationary series. Combine the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) to determine the autoregression order p and the moving average order q, and construct the ARIMA(p, d, q) model. Then, the ARIMA model is trained and its parameters are optimized using the training dataset to obtain the trained ARIMA model. Finally, the test dataset is input into the trained ARIMA model, which outputs the prediction results of the linear components.

[0067] Example 6 Based on Example 5, the nonlinear perturbation prediction of the LSTM model specifically includes the following sub-steps: Constructing a multidimensional nonlinear feature matrix: The nonlinear component N(t) is horizontally concatenated with the collected multidimensional production impact features to obtain a multidimensional input matrix containing temporal disturbance features and external impact features; Construct an LSTM network model with the following structure: input layer, first LSTM layer, dropout layer, second LSTM layer, dropout layer, fully connected layer, and output layer. The model training parameters are set as follows: batch size is 32, number of iterations is 100, the optimizer is Adam, and the loss function is mean squared error (MSE). Next, the LSTM model is trained using the training dataset, and overfitting is prevented by early stopping, resulting in a trained LSTM model. Finally, the test dataset is input into the trained LSTM model, which outputs the prediction results of the nonlinear perturbation.

[0068] Example 7 Based on Example 6, the adaptive weighted fusion based on real-time fluctuation intensity specifically includes the following sub-steps: Calculate the real-time fluctuation intensity of the work time series The sliding window method is used to calculate the current time step. The standard deviation of the working hours at each moment is used as the fluctuation intensity, and the formula is: ; in, This represents the average working hours within the sliding window. To adjust the sliding window size; Then establish the mapping relationship between fluctuation intensity and model weights: ; in, These are the output weights of the ARIMA model. The output weights of the LSTM model. and These represent the maximum and minimum values ​​of fluctuation intensity in the training dataset, respectively. Finally, the prediction results of ARIMA and LSTM are adaptively weighted and fused to obtain the final prediction result of field working hours: .

[0069] Example 8 A workshop time prediction device based on temporal feature decomposition and dynamic fusion includes a processor and a memory. The memory stores a computer program. When the computer program is loaded by the processor, it executes the application package multi-level cache distribution method based on topology perception as described in any one of Embodiments 1 to 7.

[0070] Example 9 A computer-readable storage medium storing a computer program, which, when loaded by a processor, executes the workshop time prediction method of temporal feature decomposition and dynamic fusion as described in any one of Embodiments 1 to 8.

[0071] Example 10 A workshop time prediction system based on temporal feature decomposition and dynamic fusion includes the workshop time prediction device based on temporal feature decomposition and dynamic fusion described in Example 8.

[0072] All parts not covered in this invention are the same as or can be implemented using existing technologies.

[0073] The above technical solution is only one embodiment of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the above specific embodiments of the present invention. Therefore, the methods described above are only preferred and are not restrictive.

[0074] In addition to the examples above, other embodiments may be obtained by those skilled in the art based on the above disclosure or by making modifications using knowledge or technology in related fields. The features of each embodiment may be interchanged or replaced. Modifications and changes made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting workshop time based on temporal feature decomposition and dynamic fusion, characterized in that, Including the following steps: Multi-source data acquisition and preprocessing, receiving multi-source heterogeneous data related to workshop and field working hours; The preprocessed data is then decomposed into three components of the time series of work hours. For the decomposed data, linear component prediction using the ARIMA model and nonlinear perturbation prediction using the LSTM model are employed. For the predicted data, an adaptive weighted fusion based on the real-time fluctuation intensity is performed.

2. The workshop time prediction method based on temporal feature decomposition and dynamic fusion according to claim 1, characterized in that, The multi-source heterogeneous data specifically includes: historical working hour data, field task data, and personnel skill data.

3. The workshop time prediction method based on temporal feature decomposition and dynamic fusion according to claim 1, characterized in that, The preprocessing includes missing value handling, outlier detection, and duplicate value detection. After data cleaning, numerical fields are normalized through data standardization, mapping the data to the [0, 1] interval. Feature engineering is constructed based on the preprocessed data, where the predictive features include time features, lag features, sliding window features, task features, and personnel features. Through feature engineering, an input matrix X (n×P) containing P features is finally obtained, where n is the number of samples and P is the number of features. At the same time, a target variable Y (n×1) is constructed, which is the working hour value for the corresponding date.

4. The workshop time prediction method based on temporal feature decomposition and dynamic fusion according to claim 3, characterized in that, The three-component decomposition of the working time series specifically includes the following sub-steps: The STL algorithm is used to decompose the preprocessed original field work time series Y(t) into a trend term T(t), a periodic term S(t), and a random fluctuation term R(t). The decomposition formula is as follows: Y(t) = T(t) + S(t) + R(t); Where T(t) represents the long-term trend term, reflecting the long-term linear variation law of the workshop field working hours; S(t) represents the periodic term, reflecting the periodic field law; and R(t) represents the random fluctuation term, reflecting the nonlinear random fluctuation. Then, add the trend term and the periodic term to obtain the linear stationary component L(t)=T(t)+S(t), and the random fluctuation term is taken as the nonlinear component N(t)=R(t).

5. The workshop time prediction method based on temporal feature decomposition and dynamic fusion according to claim 4, characterized in that, The linear component prediction of the ARIMA model specifically includes the following sub-steps: The stationarity of the linear stationary component L(t) is tested. The stationarity of the sequence is verified by the ADF method. If the sequence is non-stationary, it is differencing until it becomes stationary. The difference order d is then determined. Next, plot the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots of the stationary series. Combine the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) to determine the autoregression order p and the moving average order q, and construct the ARIMA(p, d, q) model. Then, the ARIMA model is trained and its parameters are optimized using the training dataset to obtain the trained ARIMA model. Finally, the test dataset is input into the trained ARIMA model, which outputs the prediction results of the linear components.

6. The workshop time prediction method based on temporal feature decomposition and dynamic fusion according to claim 5, characterized in that, The nonlinear perturbation prediction of the LSTM model specifically includes the following sub-steps: Constructing a multidimensional nonlinear feature matrix: The nonlinear component N(t) is horizontally concatenated with the collected multidimensional production impact features to obtain a multidimensional input matrix containing temporal disturbance features and external impact features; Construct an LSTM network model with the following structure: input layer, first LSTM layer, dropout layer, second LSTM layer, dropout layer, fully connected layer, and output layer. The model training parameters are set as follows: batch size is 32, number of iterations is 100, the optimizer is Adam, and the loss function is mean squared error (MSE). Next, the LSTM model is trained using the training dataset, and overfitting is prevented by early stopping, resulting in a trained LSTM model. Finally, the test dataset is input into the trained LSTM model, which outputs the prediction results of the nonlinear perturbation.

7. The workshop time prediction method based on temporal feature decomposition and dynamic fusion according to claim 6, characterized in that, The adaptive weighted fusion based on real-time fluctuation intensity specifically includes the following sub-steps: Calculate the real-time fluctuation intensity of the work time series The sliding window method is used to calculate the current time step. The standard deviation of the working hours at each moment is used as the fluctuation intensity, and the formula is: ； in, This represents the average working hours within the sliding window. To adjust the sliding window size; Then establish the mapping relationship between fluctuation intensity and model weights: ； in, These are the output weights of the ARIMA model. The output weights of the LSTM model. and These represent the maximum and minimum values ​​of fluctuation intensity in the training dataset, respectively. Finally, the prediction results of ARIMA and LSTM are adaptively weighted and fused to obtain the final prediction result of field working hours: 。 8. A workshop time prediction device based on temporal feature decomposition and dynamic fusion, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is loaded by the processor, it executes the application packet multi-level cache distribution method based on topology perception as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, A computer program is stored in a readable storage medium, which, when loaded by a processor, executes the workshop time prediction method of time-series feature decomposition and dynamic fusion as described in any one of claims 1 to 8.

10. A workshop time prediction system based on temporal feature decomposition and dynamic fusion, characterized in that, This includes a workshop time prediction device based on temporal feature decomposition and dynamic fusion as described in claim 8.

Images