Service plan scheduling method and service plan scheduling system for mining machinery

Abstract

The invention discloses a mining machinery service plan scheduling method and system, and belongs to the technical field of mining machinery, and the method comprises the steps: S1, carrying out the hierarchical data fusion, integrating the business data of a business system and the operation data of sensor equipment, and carrying out the exception processing of the data; s2, extracting spatio-temporal features to form a dynamic spatio-temporal feature fusion data set; s3, multi-task prediction is carried out by using deep learning, and fine-grained discrimination and increment correction of the effective working hours of the equipment are realized; and S4, dynamically adjusting and scheduling the service plan through data driving. The method has the advantages of integrating multi-source heterogeneous data in combination with spatial-temporal characteristics of a working place of the equipment, performing increment correction by using a deep learning algorithm, dynamically adjusting a service schedule, and ensuring good operation of the equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mining machinery, in particular to a service plan scheduling method and system for mining machinery. BACKGROUND

[0002] Mining machinery, especially heavy mining trucks and large or super-large excavators, has the characteristics of high equipment value and high reliability requirements. The operation of a mining project has the characteristics of high intensity, heavy production tasks, and complex environmental conditions. To ensure the service efficiency of the equipment and improve the accuracy of the service plan, it is an important aspect to solve the scheduling of service resources for mining.

[0003] At present, the equipment maintenance schedule for high-load mining machinery mainly focuses on the health diagnosis of local parts, ignoring the time sequence coupling relationship of multi-dimensional information in the overall operation of the equipment, and the decision basis is single. It cannot effectively dynamically adjust the service plan of the equipment in combination with the environmental characteristics of the equipment work site.

[0004] For example, Chinese patent CN109459942A discloses a large equipment active after-sales service system based on big data mining, and CN118861772A discloses an unmanned mining truck health diagnosis system based on online adaptive sampling. These systems only alarm and prompt service abnormal data, predict and warn local parts through a health score table, and remind maintenance personnel. However, due to the lack of prediction analysis of the actual operation of the equipment, they cannot effectively form a linkage with the service plan scheduling. In addition, the data used are mainly from equipment sensors, without integrating the location information and time sequence information of the equipment with the service plan, resulting in a disconnection between preventive or predictive maintenance strategies and the actual load of the equipment.

[0005] In addition, when the service resource scheduling of the equipment is in a complex environment and working condition, the integration and use of multi-source heterogeneous data is a problem that needs to be solved urgently. SUMMARY

[0006] The present application aims to solve the technical problems mentioned in the background. The first aspect provides a service plan scheduling method for mining machinery, which integrates multi-source heterogeneous data by combining the spatio-temporal characteristics of the equipment work site, uses a deep learning algorithm for incremental correction, and dynamically adjusts the service plan table to ensure the good operation of the equipment.

[0007] The second aspect provides a service plan scheduling system as the running basis for the first aspect of the service plan scheduling method for mining machinery.

[0008] According to the content of the present application, the first aspect of the present application provides a service plan scheduling method for mining machinery, which includes the following steps:

[0009] S1, obtain the device historical inspection data, device project data, business data of an integrated business system, sensor device operation data, etc. by using a CRM (customer relationship management) RESTful API v2.1 through a mine mechanical equipment coding ID, and perform data fusion in hierarchical levels and perform abnormality processing on the data;

[0010] S2, perform time series feature mining on the processed device behavior data, extract space-time features, and form a dynamic space-time feature fusion multi-action spectrum work hour feature data set;

[0011] S3, use the multi-action spectrum work hour feature data set formed in S2 as a training sample, use a deep learning framework to perform multi-task prediction, and use a sliding window mechanism to realize fine-grained discrimination and incremental correction of device effective work hours;

[0012] S4, combine the extracted device historical service execution behavior data and service process standardized data, analyze the device work hour prediction data obtained in S3 within 3-7 days, and finally form a daily scheduling plan and an abnormality checking table.

[0013] Further:

[0014] S1, obtain the device historical inspection data, device project data, business data of an integrated business system, sensor device operation data, etc. by using a CRM (customer relationship management) RESTful API v2.1 through a mine mechanical equipment coding ID, and perform data fusion in hierarchical levels and perform abnormality processing on the data;

[0015] S11, calibrate rotation speed, work hours, displacement and other data, and extract sensor behavior data of the device in the last 30 days, including device engine power-on signal, GPS return time, engine cumulative work hours, latitude and longitude, device action signal (such as sensor current, voltage), vehicle speed and other working condition fields;

[0016] S12, identify and detect abnormal data.

[0017] Further:

[0018] S12, identify and detect abnormal data.

[0019] S121, for monotonic cumulative values such as engine cumulative work hours, use an exponential smoothing algorithm to extract abnormal values and perform abnormality discrimination processing;

[0020] S122, for real-time device action signal, vehicle speed, latitude and longitude, etc. instantaneous measurement value, use Z-scores method to process abnormal value, then use adjacent value to interpolate and fill, wherein z-scores is used to measure the standardized distance of a data point from the mean, the formula is:

[0021] ,

[0022] Where x is the value of the data point, μ is the mean, σ is the standard deviation or specific scalar, the standardized distance threshold is > 2.5σ' (σ' is the modified standard deviation), and the identified abnormal value is set to null, and then linear interpolation is used adjacent value;

[0023] S123, for historical planned behavior data, establish a plan-actual mapping rule library, such as service plan time and planned standard device operating hours and actual difference greater than threshold, then mark this plan execution as an abnormal situation.

[0024] Further:

[0025] S2, the processed device behavior data is subjected to time series feature mining, spatiotemporal features are extracted, and dynamic spatiotemporal feature fusion multi-action spectrum work time feature data set is formed;

[0026] S21, according to the return time interval of the vehicle networking terminal and the set return frequency, sequence division is carried out, after determining the power-on start or shutdown of the device, the duration of different device action behaviors is calculated, the device action behavior is defined as the recording, classification and modeling of various actions of the device in the working process, including mining, transportation, unloading, rotation and other operation actions and their combinations, the selection of action sequence is adjusted according to the actual working condition and data situation;

[0027] S22, the duration (numerical variable) in each state is calculated, which constitutes the category variable and numerical variable of multi-task prediction, respectively.

[0028] Further:

[0029] S21, according to the return time interval of the vehicle networking terminal and the set return frequency, sequence division is carried out, after determining the power-on start or shutdown of the device, the duration of different device action behaviors is calculated, the device action behavior is defined as the recording, classification and modeling of various actions of the device in the working process, including mining, transportation, unloading, rotation and other operation actions and their combinations, the selection of action sequence is adjusted according to the actual working condition and data situation, define four state set S={S1, S2, S3, S4}, as the working condition-action spectrum rule set, wherein S1, S2, S3, S4 respectively represent the device state division of the action signal in different numerical intervals.

[0030] Further:

[0031] S22, calculate the duration of each state, respectively, to form a multi-task prediction of categorical variables and numerical variables;

[0032] S221, according to the GPS return time field to extract time-related features,

[0033] Time period division: T1 (6-12 hours), T2 (12-18 hours), T3 (18-24 hours), T4 (0-6 hours) (based on RTC real-time clock, UTC+8 time zone);

[0034] Weekday feature division: W ∈ {1,2,3,4,5,6,7} (integer encoding, 1 = Monday, and the others are in turn);

[0035] S222, dynamic grid encoding, based on historical trajectory points (N ≥ 1000) to calculate latitude and longitude extremes; [lat_min, lat_max] × [lng_min, lng_max] (remove ±3σ outliers), stored as SHA-256 hash value (256-bit digest);

[0036] N × N equidistant grid (N = 50-200 configurable, default 100), grid size Δlat=(lat_max-lat_min) / N, Δlng=(lng_max-lng_min) / N (precision 0.0001°, about 10 meters);

[0037] Coordinate mapping: lat_id=⌊(lat-lat_min) / Δlat⌋, lng_id=⌊(lng-lng_min) / Δlng⌋;

[0038] One-dimensional encoding: loca_id=lat_id×N + lng_id + 1 (row first, range 1-N²);

[0039] S223, extract state-grid joint features,

[0040] Through the sliding window (window length 3s, overlap 50%) to calculate V_max (Si, loca_id)=max {V | current state Si ∧ grid loca_id};

[0041] Form a certain device multi-action spectrum work time data feature set DataSets={seq1, seq2, seq3...seqi...seqn},

[0042] Wherein, seqi={dayi,weekdaysi,loca_idi,speedi,actioni,durationi},

[0043] Wherein, {(actioni,durationi)} is predicted as a multi-task variable, action is a category variable, and duration is a numerical variable.

[0044] Further:

[0045] S3, using the multi-action spectrum work time feature data set formed in S2 as a training sample, developing based on Python3.8 language, selecting PyTorch2.2 deep learning framework, and configuring CUDA 10.1 and CUDNN 7.5 environment, using the deep learning framework to perform multi-task prediction, and adopting a sliding window mechanism to realize fine-grained discrimination and incremental correction of effective work time of equipment.

[0046] S31, using a sliding window mechanism to perform incremental learning processing on the input data in S2:

[0047] The sliding window (window length L=120, step S=10) constructs a time sequence X∈R^(L×6), supports variable-length sequence input, captures long-distance dependence based on a recurrent neural network or attention mechanism sequence encoder, adopts a multi-task joint loss function, learns the action spectrum and the corresponding work time respectively, and realizes dynamic adjustment of the loss weight of each task by introducing a task precision control parameter (log_vars),

[0048] ;

[0049] Trigger condition: single-day new data > configurable threshold or precision drop > configurable threshold; retain the latest N operation cycle data (typical maintenance period of mine machinery), single-day new data > 5000 or precision drop > 5%, wherein the update mode is through weight fine-tuning (only update task head parameters, freeze shared layer); incremental learning precision decay <0.3% / month;

[0050] S32, the output content includes action spectrum probability distribution: P (S1), P (S2), P (S3), P (S4) (confidence ≥0.8), and work time prediction value: D_pred=Σ(D_i×P (S_i)) (future T period cumulative value, T=15-60 hours adjustable).

[0051] Further:

[0052] S4, combine the historical service execution behavior data of the extracted equipment, and the service process standardization data, analyze the equipment working hour prediction data within 3-7 days obtained in S3, and the specific calculation logic is as follows:

[0053] Current running time ; Last planned service content ; Last planned service time ; Next service plan standard content ; Next service plan standard time ; Lower limit time ; Upper limit time ;

[0054] Future seven-day action spectrum working hours: (Obtained through S3 step);

[0055] The judgment logic is as follows:

[0056] If a certain day in the future satisfies the following conditions, remind that something needs to be done on the t1+i day:

[0057] ;

[0058] If the following conditions are met on the first day, a pre-warning is given on the t1+i day, prompting that the upper limit of service has been exceeded and needs to be checked:

[0059] ;

[0060] If none of the above conditions are met within the next seven days , no reminders will be given,

[0061] ;

[0062] Finally, a daily scheduling plan and an abnormality checking table are formed, and the future operation trend is pushed to push or warn the service content according to the working condition.

[0063] Further:

[0064] S3, request input S2 data through Redis queue access feature extraction engine, generate time sequence feature vector and input prediction module, and the scheduling decision result is issued through REST API;

[0065] S31, the sequence encoder based on recurrent neural network or attention mechanism is Bi-LSTM / Transformer architecture.

[0066] The second aspect of the present application provides a technical solution: a service plan scheduling system, comprising a processor and a computer readable storage medium electrically connected to each other, the computer readable storage medium is adapted to store a plurality of program codes, the program codes are adapted to be loaded and run by the processor to execute the service plan scheduling method of the mining machinery provided by the first aspect;

[0067] Further, it also includes a data processing framework electrically connected with the processor and the computer readable storage medium, the data processing framework includes a data integration and processing module, a space-time feature fusion module, a multi-task collaborative prediction module and a service plan scheduling module electrically connected in sequence;

[0068] The data integration and processing module adopts a hierarchical fusion strategy to uniformly access multi-source data, and collects device operation information, customer relationship management system (CRM) and resource state three types of easy-to-buy data sources through an edge computing controller; through a lightweight ETL tool, the original data is standardized in format, and an entity resolution technology based on ontology mapping is used to perform semantic alignment on different action behaviors; through an embedded abnormal data identification and filtering unit, outlier detection algorithm is used to shield fault node noise data.

[0069] The space-time feature fusion module extracts device start-stop, load switching and other action events from the original sensor signals, uses a rule corresponding table to realize the analysis of the working condition action spectrum, extracts the running time length of each action spectrum through a difference integration mechanism and encodes it as a time sequence label, and uses space-time position encoding to fuse GPS trajectory and work area grid, and to generate a space-time action label.

[0070] The multi-task collaborative prediction module integrates the normalized feature action spectrum signal and the historical maintenance database in the input layer, uses a bidirectional recurrent neural network to construct a prediction model to analyze and capture the potential evolution law of the device action spectrum, uses a designed MultiLoss in the output layer to support multi-scale prediction, and is used for service resource core scheduling, and simultaneously uses incremental calculation based on a sliding time window to capture abnormalities in real time.

[0071] The service plan scheduling module dynamically adjusts the service plan in real time by combining the service plan data with the action spectrum work hour information provided by the collaborative prediction model.

[0072] Compared with the prior art, the present application has the following advantages:

[0073] 1. By integrating service behavior data and complex working condition work hour prediction model, the future work hour distribution under different action spectrums can be predicted, and a maintenance plan can be customized for each mining machinery, replacing the traditional static threshold distribution and temporary decision, improving the rationality of resource scheduling, and realizing more flexible on-site scheduling.

[0074] 2. By deeply integrating multi-dimensional information from connected vehicle data, including equipment status, working condition-action resource matching rules, spatial information, and historical maintenance records, this system achieves joint perception of mining machinery status and corresponding operating hours, improving the efficiency of heterogeneous data fusion and enabling fine-grained identification and prediction of effective working hours that affect equipment service plans at the micro level.

[0075] 3. The incremental learning model based on the sliding window reversely integrates the real-time working time data into the feature fusion module, and maintains the predictive performance of the model through continuous updates. This effectively solves the problem of model performance degradation caused by changes in the project environment and ensures that the system maintains continuous adaptability and accuracy in actual applications.

[0076] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 A schematic diagram of a service planning and scheduling system according to an embodiment of the present invention;

[0078] Figure 2 This is a schematic flow chart of the steps of a service planning and scheduling method for mining machinery according to an embodiment of the present invention;

[0079] Figure 3 for Figure 1 The workflow diagram of the multi-task collaborative prediction module shown in ;

[0080] Figure 4 for Figure 1 The service plan module workflow diagram is shown in . DETAILED DESCRIPTION

[0081] The present invention is described in further detail below.

[0082] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The following embodiments are used to illustrate the present invention but are not used to limit the scope of the present invention.

[0083] Example 1, combined with Figure 1 As shown, this section provides a service plan scheduling system, which serves as the operating basis of the service plan scheduling method for mining machinery in Example 2, including a data processing framework electrically connected to a processor and a computer-readable storage medium (memory), the data processing framework including a data integration and processing module, a spatiotemporal feature fusion module, a multi-task collaborative prediction module, and a service plan scheduling module electrically connected in sequence;

[0084] The data integration and processing module adopts a hierarchical fusion strategy to uniformly access multi-source data, collects device operation information through an edge computing controller, customer relationship management system (CRM) and resource state three types of easy-to-buy data sources; through a lightweight ETL tool, the original data is standardized in format and the entity resolution technology based on ontology mapping is used to perform semantic alignment on different action behaviors, through the built-in abnormal data identification and filtering unit, the outlier detection algorithm is used to shield the fault node noise data. The unified processing of batch devices is guaranteed, the data processing throughput is improved by 300%, the data set integrity is guaranteed to reach 100%, the jump data is repaired to be monotonic and non-decreasing, and the accuracy reaches 99%.

[0085] The spatio-temporal feature fusion module extracts device start-stop, load switching and other action events from the original sensor signals, uses a rule correspondence table to realize the analysis of the working condition action spectrum, extracts the running time length of each action spectrum by a difference integration mechanism and encodes it as a time sequence label, uses spatio-temporal position encoding, fuses GPS trajectory and work area grid, and fuses action spectrum to generate a spatio-temporal action label. This module converts high-frequency data through an edge device and transmits data in real time through an edge gateway. A multi-feature and multi-task time sequence feature library is constructed in the cloud to provide standardized input data for subsequent algorithm analysis.

[0086] The multi-task collaborative prediction module integrates the normalized feature action spectrum signal and the historical maintenance database in the input layer, uses a bidirectional recurrent neural network to build a prediction model to analyze and capture the potential evolution law of the device action spectrum, uses a designed MultiLoss in the output layer to support multi-scale prediction, and is used for service resource core scheduling, and simultaneously uses incremental calculation based on a sliding time window to capture abnormalities in real time.

[0087] The service plan scheduling module dynamically adjusts the service plan in real time by combining the action spectrum work hour information provided by the collaborative prediction model with the service plan data.

[0088] Embodiment two, in combination with Figure 2 , Figure 3 , Figure 4 As shown in FIG. 2, a service plan scheduling method of a mining machine comprises the following steps:

[0089] S1, the step is processed and calculated in the data integration and processing module, through the mine mechanical equipment coding ID, the historical inspection data, equipment project data, etc. of the equipment are acquired by using the data integration and processing module through CRM (customer relationship management) RESTful API v2.1, including the service plan type of the equipment, the maintenance time, the maintenance equipment working hours; through the vehicle-mounted CAN bus synchronization protocol, the speed, working hours, displacement and other data are calibrated, the sensor behavior data of the equipment in the last 30 days are extracted, including the equipment engine power-on signal, GPS return time, engine cumulative working hours, latitude and longitude, equipment action signal (such as sensor current, voltage), vehicle speed and other working condition fields. Abnormal data identification and detection are carried out. Specifically, for monotone cumulative values such as engine cumulative working hours, exponential smoothing algorithm is used to extract abnormal values and carry out abnormal discrimination processing, secondly, for real-time equipment action signal, vehicle speed, latitude and longitude and other instantaneous measurement values, Z-scores method is used to process abnormal values, and then adjacent values are used for interpolation filling, wherein z-scores is used to measure the standardized distance of a data point from the mean. The formula is wherein x is the value of the data point, μ is the mean, σ is the standard deviation or a specific scalar, the standardized distance threshold is greater than 2.5σ' (σ' is the modified standard deviation), and the identified abnormal values are set to null values, and then linear interpolation is used for adjacent values.

[0090] For historical planned behavior data, a plan-actual mapping rule library is established, for example, if the service plan time and the planned standard device running hours are greater than the threshold, the plan execution is marked as an abnormal situation.

[0091] S2, the step is processed and calculated in the space-time feature fusion module, and the processed equipment behavior data is subjected to time series feature mining.

[0092] Firstly, according to the time interval and the set return frequency of the vehicle networking terminal, the sequence is divided, after determining the power-on start or shutdown of the equipment, the duration of different equipment action behaviors is calculated, the equipment action behavior is defined as the record, classification and modeling of various actions of the equipment in the working process, including mining, transportation, unloading, rotation and other operation actions and their combinations, the selection of the action sequence is adjusted according to the actual working condition and data condition, optionally, a four-state set S={S1, S2, S3, S4} is defined as a working condition-action spectrum rule set, wherein Si represents the equipment state division when the action signal is in different numerical interval, at the same time, the duration of each state is calculated, which constitutes the category variable and the numerical variable of multi-task prediction.

[0093] Further, extract time-related features from GPS return time field, time period division: T1(6-12 hours), T2(12-18 hours), T3(18-24 hours), T4(0-6 hours) (based on RTC real-time clock, UTC+8 time zone).

[0094] Weekday feature: W ∈ {1,2,3,4,5,6,7} (integer encoding, 1=Monday, others follow suit).

[0095] Further, dynamic grid encoding is performed, and latitude and longitude extreme values [lat_min, lat_max]x[lng_min, lng_max] are calculated based on historical trajectory points (N≥1000) (excluding ±3σ outliers), and stored as SHA-256 hash values (256-bit digest). NxN equidistant grid (N=50-200 configurable, default 100), grid size Δlat=(lat_max-lat_min) / N, Δlng=(lng_max-lng_min) / N (precision 0.0001°, about 10 meters). Coordinate mapping: lat_id=⌊(lat-lat_min) / Δlat⌋, lng_id=⌊(lng-lng_min) / Δlng⌋.

[0096] One-dimensional encoding: loca_id=lat_idxN + lng_id + 1 (row-major, range 1-N²).

[0097] Further, state-grid joint features are calculated by sliding window (window length 3s, overlap 50%) V_max(Si, loca_id)=max {V | current state Si ∧ grid loca_id}.

[0098] Finally, a multi-action profile work hour dataset DataSets={seq1, seq2, seq3...seqi...seqn} is formed for a certain device, where seqi={dayi, weekdaysi, loca_idi, speedi, actioni, durationi}, and {(actioni, durationi)} are used as multi-task variables for prediction, action is a categorical variable, and duration is a numerical variable. The entire process is encoded in the cloud to realize multi-task work condition calculation

[0099] S3, the step is processed and calculated in the multi-task collaborative prediction module, and the feature dataset formed in step S2 is used as a training sample,

[0100] This module is developed based on Python 3.8 language, selects PyTorch 2.2 deep learning framework, and configures CUDA 10.1 and CUDNN 7.5 environment to fully utilize the computing power of GPU and improve the efficiency of model training and prediction. Processing: sliding window (window length L=120, step S=10) constructs time series X∈R^(L×6), supports variable-length sequence input, sequence encoder based on recurrent neural network or attention mechanism (optional Bi-LSTM / Transformer architecture) to capture long-range dependencies, this module designs a multi-task joint loss function to learn the action spectrum and the corresponding work hours respectively, and realizes the dynamic adjustment of the loss weight of each task by introducing the task precision control parameter (log_vars).

[0101]

[0102] This module uses sliding window mechanism to realize incremental learning, trigger condition: single-day new data > configurable threshold or precision drop > configurable threshold. Retain the last N job cycle data (typical maintenance period of mining machinery), single-day new data > 5000 or precision drop > 5%, wherein the update method is through weight fine-tuning (only update task head parameters, freeze shared layer). Incremental learning precision decay <0.3% / month.

[0103] Finally, the output content includes action spectrum probability distribution: P (S1), P (S2), P (S3), P (S4) (confidence ≥0.8), work hour prediction value: D_pred=Σ(D_i×P (S_i)) (future T period cumulative value, T=15-60 hours adjustable). Request data accesses the feature extraction engine through the Redis queue, generates a time series feature vector, and inputs the prediction module. The scheduling decision result is delivered through the REST API, and the interface specification is: REST API (complies with ISO 29481-1:2010 mechanical information model).

[0104] S4, this step is processed and calculated in the service plan scheduling module, such as Figure 4 The prediction result is pushed to the relevant business system to complete the scheduling and reminding of the service plan.

[0105] Further, an embodiment of a service plan scheduling scheme is provided.

[0106] Combined with the extracted historical service execution behavior data of a single equipment and the standardized data of the service process, the 3-7 day equipment work hour prediction data obtained from S3 is analyzed, and the specific calculation logic is as follows:

[0107] Current running time Last planned service content Last planned service time Next service plan standard content Next service plan standard time Lower limit time Upper limit time Future seven-day action spectrum working hours: Acquired through the S3 step

[0108] Judgment logic:

[0109] If the following conditions are met on a certain day in the future , remind that a certain service needs to be done on the t1+i day:

[0110] ;

[0111] If the following conditions are met on the first day, give a warning on the t1+i day, prompt that the upper limit of service has been exceeded and needs to be investigated:

[0112] ;

[0113] If the following conditions are not met within the next seven days , do not remind:

[0114] ;

[0115] Finally, the following Table 1 is formed, and the future operation trend is pushed to push and warn the service content according to the working condition.

[0116] When predicting that the equipment will perform a certain service, a report can be pushed in the form of a report. The work unit cooperates according to the equipment maintenance manual, intelligently schedules appropriate maintenance personnel, automatically allocates maintenance tasks, and notifies the maintenance personnel. Through the daily maintenance standard manual, real-time pricing plans are pushed. The service resource scheduling error is reduced by 35%-42%.

[0117] Table 1 Service plan table

[0118] Device code Device status Current running time Last service content Expected service time Next service content Service expected time A12 Normal operation 2024-10-07 Check item 1 Expected maintenance in 5 days Check item 2 2024-10-12

[0119] From the above, the advantages of the present application are:

[0120] 1. The multi-task prediction scheduling framework of spatio-temporal feature fusion is first proposed, which breaks through the limitations of traditional single-dimensional modeling and static scheduling, solves the technical problems of multi-dimensional data processing, real-time dynamic scheduling and model continuous adaptability of mine machinery under high load and multiple working conditions.

[0121] 2. The application innovatively adopts a space-time multi-dimensional feature fusion technology, constructs a complete device operation portrait by deeply extracting and integrating spatial trajectory information, action spectrum features and time dimension features of the mine equipment, and greatly improves the model's representation ability for time sequence features, realizing high-precision identification of the corresponding action spectrum working hours of the service maintenance of the device.

[0122] 3. The utilization efficiency and response rate of service resources are improved by dynamically adjusting the service plan scheduling framework.

[0123] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A service planning and scheduling method for mining machinery, characterized in that: The following steps are involved: S1. Using the mining machinery equipment ID, use the CRM (Customer Relationship Management) RESTful API v2.1 to obtain the equipment's historical scheduled inspection data, equipment project data, business data from integrated business systems, sensor equipment operation data, etc., and perform hierarchical data integration and data exception processing. S2. Mining the time series features of the processed equipment behavior data to extract spatiotemporal features and form a dynamic spatiotemporal feature fusion multi-action spectrum working time feature dataset; S3, using the multi-action spectrum working time feature dataset formed in S2 as training samples, using the deep learning framework to perform multi-task prediction, and using the sliding window mechanism to achieve fine-grained discrimination and incremental correction of the equipment's effective working time; S4: Combine the extracted historical service execution behavior data of the equipment and the standardized service process data, and analyze the 3-7 day equipment working hour forecast data obtained in S3 to finally form a daily shift schedule and anomaly troubleshooting table.

2. The service planning and scheduling method for mining machinery according to claim 1, characterized in that: S1. Using the mining machinery equipment ID, use the CRM (Customer Relationship Management) RESTful API v2.1 to obtain the equipment's historical scheduled inspection data, equipment project data, business data from integrated business systems, sensor equipment operation data, etc., and perform hierarchical data integration and data exception processing. S11. Calibrate data such as speed, working hours, and displacement, and extract sensor behavior data for the device over the past 30 days, including engine power-on signals, GPS return time, engine cumulative working hours, longitude and latitude, device action signals (such as sensor current and voltage), vehicle speed, and other operating condition fields. S12. Identify and detect abnormal data.

3. The service planning and scheduling method for mining machinery according to claim 2, characterized in that: S12, identifying and detecting abnormal data; S121. For monotonically accumulated values ​​such as the accumulated engine working hours, an exponential smoothing algorithm is used to extract abnormal values ​​and perform abnormality discrimination processing; S122. For instantaneous measurement values ​​such as real-time device action signals, vehicle speed, and latitude and longitude, use the Z-score method to process outliers and then interpolate and fill in the gaps with neighboring values. The Z-score is used to measure the standardized distance a data point deviates from the mean. The formula is: ， Where x is the value of the data point, μ is the mean, σ is the standard deviation or a specific scalar, the standardized distance threshold is > 2.5σ' (σ' is the corrected standard deviation), the identified outliers are set to null values, and then linearly interpolated with neighboring values; S123. For historical plan behavior data, a plan-actual mapping rule base is established. If the difference between the service plan time and the equipment operating hours specified in the plan standard and the actual hours is greater than a threshold, the plan execution is marked as an abnormal situation.

4. The service planning and scheduling method for mining machinery according to claim 1, characterized in that: S2. Mining the time series features of the processed equipment behavior data to extract spatiotemporal features and form a dynamic spatiotemporal feature fusion multi-action spectrum working time feature dataset; S21. Sequence the data based on the return intervals of the connected vehicle terminals and the set return frequency. After determining whether the equipment is powered on or off, calculate the duration of different equipment actions. Equipment action behavior is defined as the recording, classification, and modeling of various actions of the equipment during operation, including excavation, transportation, unloading, rotation, and other operating actions, as well as their combinations. The selection of action sequences is adjusted based on actual working conditions and data. S22. Calculate the duration (numeric variable) of each state to form the categorical variable and numerical variable of multi-task prediction.

5. The service planning and scheduling method for mining machinery according to claim 4, characterized in that: S21. Sequence division is performed based on the feedback time interval of the Internet of Vehicles terminal and the set feedback frequency. After determining whether the equipment is powered on or shut down, the duration of different equipment action behaviors is calculated. Equipment action behavior is defined as the recording, classification and modeling of various actions of the equipment during the working process, including excavation, transportation, unloading, rotation and other working actions and their combinations. The selection of its action sequence is adjusted according to the actual working conditions and data conditions. The four-element state set S={S1,S2,S3,S4} is defined as the working condition-action spectrum rule set, where S1, S2, S3, and S4 respectively represent the equipment state division when the action signal is in different numerical ranges.

6. The service planning and scheduling method for mining machinery according to claim 4, characterized in that: S22. Calculate the duration of each state to form the categorical variable and numerical variable for multi-task prediction respectively; S221, extracting time-related features based on the GPS return time field, Time period division: T1 (6-12 o'clock), T2 (12-18 o'clock), T3 (18-24 o'clock), T4 (0-6 o'clock) (based on RTC real-time clock, UTC+8 time zone); Week feature partitioning: W∈{1,2,3,4,5,6,7} (integer encoding, 1 = Monday, and so on); S222. Perform dynamic grid coding and calculate the extreme longitude and latitude values ​​based on historical trajectory points (N ≥ 1000); [lat_min, lat_max] × [lng_min, lng_max] (excluding ±3σ outliers), and store them as SHA-256 hash values ​​(256-bit digest). N×N equidistant grid (N=50-200 configurable, default 100), grid size Δlat=(lat_max-lat_min) / N, Δlng=(lng_max-lng_min) / N (accuracy 0.0001°, about 10 meters); Coordinate mapping: lat_id = ⌊(lat-lat_min) / Δlat⌋, lng_id = ⌊(lng-lng_min) / Δlng⌋; One-dimensional encoding: loca_id = lat_id × N + lng_id + 1 (row-first, range 1-N²); S223, extracting state-grid joint features, Calculate V_max (Si,loca_id)=max {V | current state Si ∧ grid loca_id} through a sliding window (window length 3s, overlap 50%); Form a feature set of multi-action spectrum working time data for a certain equipment: DataSets={seq1,seq2,seq3...seqi...seqn}, Among them, seqi={dayi,weekdaysi,loca_idi,speedi,actioni,durationi}, Among them, {(actioni,durationi)} is predicted as a multi-task variable, action is a categorical variable, and duration is a numerical variable.

7. The service planning and scheduling method for mining machinery according to claim 1, characterized in that: S3 uses the multi-action spectrum working time feature dataset generated in S2 as training samples. Development is conducted in Python 3.8, using the PyTorch 2.2 deep learning framework, and configuring the CUDA 10.1 and CUDNN 7.5 environments. Multi-task prediction is performed using the deep learning framework, and a sliding window mechanism is employed to achieve fine-grained discrimination and incremental correction of equipment effective working hours. S31, use the sliding window mechanism to perform incremental learning on the data in S2: A sliding window (window length L = 120, step size S = 10) is used to construct a time series sequence X∈R^(L×6), which supports variable-length sequence input. A sequence encoder based on a recurrent neural network or attention mechanism captures long-distance dependencies. A multi-task joint loss function is used to learn the action spectrum and the corresponding working time separately. By introducing a task accuracy control parameter (log_vars), the loss weight of each task can be dynamically adjusted. ； Trigger conditions: Daily new data > configurable threshold or accuracy drop > configurable threshold; retain data from the most recent N operating cycles (typical maintenance cycles for mining machinery); daily new data > 5,000 items or accuracy drop > 5%. Updates are made through weight fine-tuning (updating only task header parameters and freezing shared layers); incremental learning accuracy decay < 0.3% / month; S32. Output includes the probability distribution of motion spectrum: P (S1), P (S2), P (S3), P (S4) (confidence ≥ 0.8), and the predicted working hours: D_pred = Σ(D_i × P (S_i)) (accumulated value in the future T period, T = 15-60 hours, adjustable).

8. The service planning and scheduling method for mining machinery according to claim 1, characterized in that: S4. Combine the extracted historical service execution behavior data of the equipment and the standardized service process data, and analyze the 3-7-day equipment working hour forecast data obtained in S3. The specific calculation logic is as follows: Current running time ;Last planned service content ; Last scheduled service time ; Standard content of the next service plan ; Next service plan standard time ; Lower limit time Upper limit time ; Working hours by action score for the next seven days: (obtained via the S3 step); The judgment logic is as follows: If one day in the future If the following conditions are met, a reminder to perform a certain service will be given on day t1+i: ； If the following conditions are met on the first day, an alert will be issued on day t1+i, indicating that the service limit has been exceeded and that an investigation is required: ； If it is not reached within the next seven days , then no reminder: ； Finally, a daily shift schedule and anomaly troubleshooting table are formed, and future operation trends are pushed to provide push and early warning of service content based on working conditions.

9. The service planning and scheduling method for mining machinery according to claim 7, characterized in that: S3: The requested S2 data is connected to the feature extraction engine through the Redis queue, and the time series feature vector is generated and input into the prediction module. The scheduling decision result is issued through the REST API; S31. The sequence encoder based on recurrent neural network or attention mechanism refers to the Bi-LSTM / Transformer architecture.

10. A service planning and scheduling system, characterized by: The method comprises a processor and a computer-readable storage medium electrically connected to each other, the computer-readable storage medium being suitable for storing a plurality of program codes, the program codes being suitable for being loaded and run by the processor to execute the service plan scheduling method for mining machinery according to any one of claims 1 to 9; It also includes a data processing framework electrically connected to the processor and the computer-readable storage medium, and the data processing framework includes a data integration and processing module, a spatiotemporal feature fusion module, a multi-task collaborative prediction module and a service plan scheduling module electrically connected in sequence.

Images