A multi-task prediction method for target objects

Abstract

This application provides a multi-task prediction method for a target object. It establishes prediction units according to the spatial evaluation scale of the target area and observation periods according to the periodic state changes of the target object. Multi-source data from different acquisition channels are grouped into unit data groups based on spatial and temporal attribution. Data quality identifiers are generated based on quality evaluation information, and the object's stage state is determined based on multiple unit data groups within the same prediction unit. First and second task adaptation parameters are generated by combining the object's stage state and data quality identifiers, forming first and second prediction inputs, resulting in first and second prediction results. In response to newly added evaluation data, prediction deviations are determined, and deviations are located and corrected based on spatial, temporal, and correlational relationships.

Description

Technical Field

[0001] This application relates to the field of data processing and prediction technology, and in particular to a multi-task prediction method for a target object. Background Technology

[0002] With the development of data acquisition, device networking, and model prediction technologies, the method of predicting target objects based on multi-source data has been applied to technical scenarios such as agricultural production, industrial manufacturing, warehousing and logistics, energy systems, and equipment operation monitoring. In these scenarios, it is usually necessary to collect data from different sources and output prediction results for different prediction tasks of target objects based on this data.

[0003] In related technologies, data from different acquisition channels are typically preprocessed, feature extracted, and stitched or fused before being input into a prediction model to obtain prediction results for the target object under one or more prediction tasks. For example, in agricultural scenarios, crop yield or quality can be predicted based on remote sensing observation data, ground-based data, meteorological and soil data, agricultural management data, and measured sample data; in industrial production scenarios, product output or product quality can be predicted based on visual inspection data, equipment operation data, process parameters, environmental data, and final inspection and evaluation data.

[0004] However, in practical applications, multi-source data often exhibit differences in data origin, collection scale, collection time, data completeness, anomalies, and sample representativeness. Furthermore, the actual state of the target object varies across different regions, processing units, and time periods, and the data dependencies of different prediction tasks also differ. If a uniform data processing or fusion method is used to form the prediction input, it can easily lead to a mismatch between the prediction input and the actual state of the target object and the requirements of the corresponding prediction task, thus affecting the accuracy and reliability of the prediction results.

[0005] Furthermore, when subsequent evaluation data is obtained and deviations in the prediction results are found, the relevant technologies usually focus on making overall corrections to the prediction results or prediction models, making it difficult to effectively analyze the causes of the prediction deviations, resulting in a lack of targeted corrections in subsequent corrections. Summary of the Invention

[0006] This application provides a multi-task prediction method for target objects to solve one of the aforementioned technical problems.

[0007] The technical solution adopted in this application is as follows: This application provides a multi-task prediction method for a target object, characterized by comprising: S100 establishes multiple prediction units according to the spatial evaluation scale of the target area, and establishes multiple observation periods according to the periodic state change process of the target object; S200: Acquire multi-source data from different acquisition channels, and classify the multi-source data into corresponding prediction units and corresponding observation periods according to the spatial and temporal attribution of the multi-source data to form multiple unit data groups; wherein, each unit data group corresponds to a prediction unit and an observation period, and the multi-source data includes at least two types of data used to characterize the state of the target object, the environment in which the target object is located, the operation process of the target object, and the evaluation results of the target object; S300, for the multi-source data in each unit data group, a corresponding data quality identifier is generated based on the quality evaluation information of the multi-source data in the corresponding prediction unit and the corresponding observation period; wherein, the data quality identifier is used to characterize the availability of the multi-source data in the corresponding prediction unit and the corresponding observation period for participating in subsequent prediction processing; S400, based on multiple unit data groups corresponding to the same prediction unit, feature extraction and correlation analysis are performed on the state data of the target object under multiple observation periods to obtain the object stage state of the prediction unit under the corresponding observation period. S500, based on the object stage status of the current prediction unit in the current observation period and the data quality identifier of the multi-source data in the current unit data group, generate a first task adaptation parameter for the first prediction task and a second task adaptation parameter for the second prediction task respectively; wherein, the first task adaptation parameter is used to characterize the effectiveness of the corresponding multi-source data participating in the first prediction task, and the second task adaptation parameter is used to characterize the effectiveness of the corresponding multi-source data participating in the second prediction task. S600: Based on the first task adaptation parameters, perform matching and fusion processing on the multi-source data in the current unit data group for the first prediction task to form a first prediction input; based on the second task adaptation parameters, perform matching and fusion processing on the multi-source data in the current unit data group for the second prediction task to form a second prediction input; and determine the correlation between the first prediction input and the second prediction input and the multi-source data and task adaptation parameters used to form the corresponding prediction inputs. S700, input the first prediction input and the second prediction input into the prediction model to obtain the first prediction result and the second prediction result; S800, in response to newly added evaluation data, compare the newly added evaluation data with the corresponding first prediction result or second prediction result to obtain the prediction deviation; based on the spatial and temporal attribution of the newly added evaluation data and the correlation relationship, locate the target prediction unit, target observation period, corresponding object stage state and multi-source data involved in generating the corresponding prediction result corresponding to the prediction deviation, and perform correction processing on the task adaptation parameters and / or prediction processing parameters under the corresponding prediction task based on the location results.

[0008] According to one embodiment of this application, in step S100, multiple prediction units are established according to the spatial evaluation scale of the target area, including: Obtain the spatial boundary information and spatial evaluation requirements of the target area; Based on the spatial boundary information and the spatial evaluation requirements, the target area is divided into multiple regional units that can respectively perform data collection, state judgment, prediction processing and feedback correction, and the regional units are used as the prediction units. In step S100, multiple observation periods are established according to the periodic state change process of the target object, including: The data processing interval is determined based on at least one of the following: the cycle start time, cycle end time, multi-source data acquisition frequency, and runtime event occurrence time of the target object. The multiple observation periods are established according to the data processing intervals.

[0009] According to one embodiment of this application, in step S200, the multi-source data includes at least two types of data selected from detection data, status data, environmental data, operational data, historical data, correlation data, and evaluation data. The detection data is used to characterize the detection results of the target object; The state data is used to characterize the state changes of the target object under the corresponding prediction unit and the corresponding observation period; The environmental data is used to characterize the environmental state of the environment in which the target object is located; The operational data is used to characterize operational events or processing events that act on the target object; The historical data is used to characterize the state changes or prediction results of the target object during the historical observation period; The associated data is used to characterize the relationship between the target object and the prediction task; The evaluation data is used as a basis for feedback correction of the first or second prediction result.

[0010] According to one embodiment of this application, in step S200, the multi-source data is assigned to the corresponding prediction unit and the corresponding observation period based on the spatial and temporal attribution of the multi-source data, including: When the multi-source data is planar data, the corresponding prediction unit is determined according to the overlap relationship between the coverage area of ​​the planar data and the prediction unit. When the multi-source data is point data, the corresponding prediction unit is determined according to the positional relationship between the acquisition location, detection location or sample location of the point data and the prediction unit; When the multi-source data is event data, the corresponding prediction unit is determined according to the scope of effect, operating scope or correlation scope of the event data; The corresponding observation period is determined based on the acquisition time, detection time, occurrence time, evaluation time, or correlation time of the multi-source data.

[0011] According to an embodiment of this application, in step S300, the quality evaluation information includes at least two of the following: integrity evaluation parameters, timeliness evaluation parameters, spatial matching evaluation parameters, anomaly evaluation parameters, and sample support evaluation parameters. The integrity evaluation parameter is used to characterize the data integrity of the multi-source data in the corresponding prediction unit and the corresponding observation period. The timeliness evaluation parameter is used to characterize the degree of time matching between the multi-source data and the corresponding observation period; The spatial matching evaluation parameter is used to characterize the degree of spatial matching between the multi-source data and the corresponding prediction unit; The anomaly evaluation parameters are used to characterize the degree of anomaly in the multi-source data; The sample support evaluation parameters are used to characterize the degree of support that the evaluation data or sample data provides for the corresponding prediction unit and the corresponding observation period. The data quality identifier is generated based on at least two of the quality evaluation information items and their corresponding weight parameters.

[0012] According to one embodiment of this application, in step S300, the data quality identifier includes a data quality score and a data quality level determined based on the data quality score; The data quality score Q is determined in the following manner: Q=wC×C+wT×T+wS×S+wA×A+wN×N; Wherein, C is the integrity evaluation parameter, T is the timeliness evaluation parameter, S is the spatial matching evaluation parameter, A is the anomaly evaluation parameter, and N is the sample support evaluation parameter; wC, wT, wS, wA, and wN are the weight parameters for the corresponding evaluation parameters; The sum of the weight parameters involved in the calculation is 1, and the weight parameters corresponding to the evaluation parameters not involved in the calculation are 0.

[0013] According to one embodiment of this application, in step S400, based on multiple unit data groups corresponding to the same prediction unit, feature extraction and correlation analysis are performed on the state data of the target object under multiple observation periods to obtain the object stage state of the prediction unit under the corresponding observation period, including: Multiple data sets corresponding to the same prediction unit are arranged in chronological order of the observation periods; Extract the state feature parameters of the target object from the arranged multiple cell data groups; Based on the changing relationship of the state characteristic parameters over multiple observation periods, the state change trend is obtained; Based on the state change trend, and in conjunction with environmental assessment parameters formed from environmental data and / or operational event parameters formed from operational data, the object stage state of the prediction unit under the corresponding observation period is determined.

[0014] According to one embodiment of this application, step S500 involves generating first task adaptation parameters for a first prediction task and second task adaptation parameters for a second prediction task, including: Based on the object stage state, determine the first stage matching result of the multi-source data in the current unit data group relative to the first prediction task, and the second stage matching result relative to the second prediction task. Based on the data quality identifier, determine the quality evaluation result of the multi-source data in the current prediction unit and the current observation period; Based on the first-stage matching results and the quality evaluation results, the first task adaptation parameters are generated. Based on the matching results of the second stage and the quality evaluation results, the second task adaptation parameters are generated; Among them, the first task adaptation parameters and the second task adaptation parameters corresponding to the same multi-source data can be different.

[0015] In some embodiments, the first task adaptation parameters can be generated based on the first-stage matching results, data quality identifiers, and first task association parameters; the second task adaptation parameters can be generated based on the second-stage matching results, data quality identifiers, and second task association parameters. The first and second task association parameters are used to characterize the historical or rule-based association degree between the corresponding multi-source data and the corresponding prediction task.

[0016] According to one embodiment of this application, step S600 involves performing matching and fusion processing on the multi-source data in the current unit data group, oriented towards a first prediction task or a second prediction task, including: When the data quality identifier of any multi-source data does not meet the preset judgment rules, reduce the participation of that multi-source data in the corresponding prediction input, or remove that multi-source data from the corresponding prediction input. When there are other multi-source data in the current unit data group that are related to the same prediction task and whose data quality identifiers meet the preset judgment rules, the participation of the other multi-source data in the corresponding prediction input is increased. The increase in the participation level of the other multi-source data is determined based on the task adaptation parameters of the multi-source data whose participation level has been reduced or which has been removed, as well as the task adaptation parameters of the other multi-source data relative to the same prediction task.

[0017] In some embodiments, when any multi-source data has its participation level reduced or is removed, the release participation level corresponding to that multi-source data is determined, and the release participation level is allocated to the other multi-source data according to the task adaptation parameter ratio of other multi-source data under the same prediction task.

[0018] According to one embodiment of this application, in step S800, the task adaptation parameters and / or prediction processing parameters under the corresponding prediction task are corrected based on the positioning results, including: When the newly added evaluation data corresponds to the first prediction task, the newly added evaluation data is compared with the first prediction result to obtain the first prediction deviation, and the first task adaptation parameters of the multi-source data that participated in generating the first prediction result are adjusted according to the first prediction deviation in the target prediction unit, the target observation period and the corresponding object stage state. When the newly added evaluation data corresponds to the second prediction task, the newly added evaluation data is compared with the second prediction result to obtain the second prediction deviation, and the second task adaptation parameters of the multi-source data involved in generating the second prediction result are adjusted according to the second prediction deviation in the target prediction unit, the target observation period and the corresponding object stage state. The adjustment amount of the first task adaptation parameter or the second task adaptation parameter is determined based on the corresponding prediction deviation, the data quality identifier of the corresponding multi-source data, and the degree of participation of the corresponding multi-source data in the prediction input.

[0019] Due to the adoption of the above technical solution, the beneficial effects achieved by this application are as follows: This application establishes multiple prediction units based on the spatial evaluation scale of the target area and multiple observation periods based on the periodic state changes of the target object. This ensures that subsequent multi-source data no longer directly participates in prediction processing based on the original acquisition range or time, but is instead transformed to a unified spatial and temporal processing benchmark. Consequently, data from different acquisition channels, spatial scales, and acquisition frequencies can be organized around the same prediction unit and the same observation period, providing a consistent data processing object for subsequent data quality evaluation, object stage state determination, task adaptation parameter generation, and prediction deviation localization, reducing prediction input confusion caused by inconsistencies in spatial and temporal attribution.

[0020] This application categorizes multi-source data into corresponding prediction units and observation periods based on their spatial and temporal assignments, forming multiple unit data groups. Since each unit data group corresponds to one prediction unit and one observation period, these groupings can characterize the state, environment, operational process, or evaluation results of the target object within a specific spatial range and observation period. Compared to directly aggregating multi-source data and inputting it into the prediction model, this approach enables subsequent prediction processing to be performed based on data units with clearly defined spatial and temporal assignments.

[0021] This application generates corresponding data quality identifiers for multi-source data within each data unit, based on the quality evaluation information of the multi-source data in the corresponding prediction unit and observation period. Since different multi-source data may differ in terms of completeness, timeliness, spatial matching, anomalies, or sample support, the data quality identifier can characterize the usability of the multi-source data in the current prediction unit and observation period for subsequent prediction processing, thus providing a basis for adjusting the degree of data participation.

[0022] This application uses multiple unit data groups corresponding to the same prediction unit to perform feature extraction and correlation analysis on the state data of the target object under multiple observation periods, thereby obtaining the object stage state of the prediction unit under the corresponding observation period. Since the object stage state is determined based on the data changes of the same prediction unit under multiple observation periods, rather than solely based on a uniform time rule, it can better reflect the actual state change process of the target object within the corresponding prediction unit.

[0023] This application generates first task adaptation parameters for the first prediction task and second task adaptation parameters for the second prediction task, based on the current object stage status of the current prediction unit during the current observation period and the data quality identifiers of the multi-source data in the current unit's data group. Since different prediction tasks have different requirements for multi-source data, by generating task adaptation parameters separately, the same multi-source data can be expressed with different degrees of effectiveness under different prediction tasks.

[0024] This application generates a first predictive input based on a first task adaptation parameter and a second predictive input based on a second task adaptation parameter. Since the first and second predictive inputs are generated under the control of different task adaptation parameters, they can reflect differences in data participation, data fusion method, and data matching objects, thereby reducing the interference of data with low relevance or low availability to the corresponding predictive task on the predictive input.

[0025] This application establishes the correlation between the first and second prediction inputs and the multi-source data and task adaptation parameters used to form the corresponding prediction inputs. By establishing this correlation, the multi-source data and task adaptation parameters upon which the prediction results depend can be traced after the prediction results are generated. This allows the prediction bias to be further correlated with the data sources, data quality identifiers, and task adaptation parameters involved in generating the prediction results when new evaluation data appears.

[0026] This application inputs a first prediction input and a second prediction input into a prediction model to obtain a first prediction result and a second prediction result. Since the first prediction input and the second prediction input have undergone matching and fusion processing for different prediction tasks, the input received by the prediction model can correspond to different prediction tasks, thus avoiding multiple prediction tasks from sharing the same input structure indiscriminately.

[0027] In response to newly added evaluation data, this application compares the newly added evaluation data with the corresponding first or second prediction result to obtain the prediction deviation. Based on the spatial and temporal attribution and correlation of the newly added evaluation data, it locates the target prediction unit, target observation period, corresponding object stage state, and multi-source data involved in generating the corresponding prediction result that correspond to the prediction deviation. Therefore, subsequent corrections can be performed on the local data processing chain that caused the deviation.

[0028] This application corrects the task adaptation parameters and / or prediction processing parameters for the corresponding prediction task based on the positioning results. Since the task adaptation parameters are used to characterize the effectiveness of multi-source data participating in the corresponding prediction task, when new evaluation data indicates that there is a deviation in the prediction results, the participation relationship of relevant multi-source data in the corresponding prediction task can be adjusted according to the deviation positioning results, so that the formation process of prediction input quantities in subsequent prediction units, observation periods and object stage states can be corrected. Attached Figure Description

[0029] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a multi-task prediction method for a target object provided in an embodiment of this application. Detailed Implementation

[0030] To enable those skilled in the art to better understand the technical solutions of this application, the implementation methods of this application are described below in conjunction with specific application scenarios. It should be understood that the following embodiments are only used to illustrate optional implementation methods of this application and do not constitute a limitation on the scope of protection of this application. Where there is no conflict, the technical features in the following embodiments can be combined with each other.

[0031] In this application, the target area refers to the spatial range within which predictive processing needs to be performed on the target object. The target object refers to an object within the target area that undergoes a phased state change process and requires predictive processing for at least two predictive tasks. Multi-source data refers to data from different acquisition channels, data types, or processing stages, and may include at least two of the following: detection data, status data, environmental data, operational data, historical data, correlated data, and evaluation data. The object's phase state refers to the state stage of the target object within the corresponding predictive unit and observation period. The first and second predictive tasks refer to data processing tasks corresponding to different predictive objectives. Newly added evaluation data refers to data obtained after the prediction results are generated and can be used for comparison with the corresponding prediction results.

[0032] Example 1 In this embodiment, the target area is an agricultural planting area. This target area can be a planting area, field, farm, management zone, or gridded planting area. The target object is the target crop, which can be wheat, corn, rice, soybeans, fruits and vegetables, or other crops with a cyclical growth process.

[0033] The first forecasting task is to predict crop yield, and the second forecasting task is to predict crop quality. The first forecast result can be yield per unit area, regional yield, yield grade, or yield risk status. The second forecast result can be quality index value, quality grade, or quality risk status.

[0034] In this embodiment, multi-source data may include detection data, status data, environmental data, operational data, historical data, correlated data, and evaluation data. Detection data may include remote sensing detection data, UAV detection data, ground image detection data, and sample detection data. Status data may include crop canopy status, crop growth status, late-stage decline status, and pre-harvest risk status. Environmental data may include temperature, moisture, light, soil, and meteorological data. Operational data may include agricultural event data such as sowing, fertilization, irrigation, plant protection, weeding, and harvesting. Historical data may include historical yield, historical quality, and historical status data. Correlated data may include the correlation between different data types and crop yield prediction tasks and crop quality prediction tasks. Evaluation data may include measured yield data, measured quality data, sample detection data, harvest measurement data, and quality grade evaluation data.

[0035] The target stage can be the crop's growth stage, such as seedling stage, vegetative growth stage, flowering stage, grain-filling stage, maturity stage, or pre-harvest stage. New evaluation data can include measured yield data, measured quality data, sample test data, harvest measurement data, or quality grade evaluation data.

[0036] like Figure 1 As shown, in this embodiment, multiple prediction units are established according to the spatial evaluation scale of the target area, and multiple observation periods are established according to the periodic state change process of the target crop.

[0037] Specifically, spatial boundary information and spatial evaluation requirements for the target area are obtained. Spatial boundary information may include planting boundaries, field boundaries, management zone boundaries, remote sensing identification boundaries, or boundaries of areas with different growth patterns. Spatial evaluation requirements can be used to determine the spatial granularity of subsequent prediction outputs and feedback corrections.

[0038] In some embodiments, when the target area has clearly defined field boundaries and yield and quality prediction results need to be output on a field-by-field basis, each field is considered a prediction unit. When the target area is large and requires refined prediction, the target area can be divided into multiple grid units. When different agricultural management methods exist within the target area, management zones can be used as prediction units. When detection data or status data indicates differences in growth within the target area, prediction units can be established based on areas of growth difference. Prediction units can be field units, grid units, management zone units, or units representing areas of growth difference.

[0039] Furthermore, multiple observation periods are established based on the target crop's cycle start time, cycle end time, multi-source data acquisition frequency, and operational event occurrence time. The cycle start time can be the sowing time, emergence time, or transplanting time; the cycle end time can be the maturity time, harvest time, or final evaluation time. The operational event occurrence time can be the time of occurrence of agricultural events such as fertilization, irrigation, plant protection, weeding, or harvesting. The observation periods can be divided by day, week, ten-day period, month, or crop growth stage range, and can also be adjusted according to the data acquisition frequency and agricultural event occurrence time.

[0040] Thus, the target area is divided into multiple spatial units capable of carrying out data collection, state judgment, prediction processing, and feedback correction, and the growth process of the target crop is divided into multiple observation periods capable of carrying out time attribution judgment and state change analysis.

[0041] In this embodiment, multi-source data from different acquisition channels are acquired, and the multi-source data are assigned to the corresponding prediction units and the corresponding observation periods according to the spatial and temporal affiliations of the multi-source data, forming multiple unit data groups.

[0042] Specifically, for remote sensing or UAV detection data, spatial attribution can be determined based on the overlap between the image coverage area and the prediction unit, and temporal attribution can be determined based on the image acquisition time. For ground image detection data or sample detection data, spatial attribution can be determined based on the positional relationship between the acquisition location, detection location, or sample location and the prediction unit, and temporal attribution can be determined based on the acquisition time, detection time, or sample time. For environmental data, spatial attribution can be determined based on the sensor location, representative area, or data coverage area, and temporal attribution can be determined based on the acquisition time. For agricultural event data, spatial attribution can be determined based on the scope of the agricultural event, operational trajectory, or management scope, and temporal attribution can be determined based on the event occurrence time. For evaluation data, spatial attribution can be determined based on the sample location, measurement area, or detection area, and temporal attribution can be determined based on the sampling time, detection time, harvest time, or evaluation time.

[0043] In some embodiments, multi-source data can be categorized into areal data, point data, and event data. Areal data can be remote sensing imagery or UAV imagery, and its spatial attribution can be determined based on the overlap between the coverage area and the prediction unit. Point data can be soil sensor data, sample detection data, or ground image acquisition data, and its spatial attribution can be determined based on the acquisition location, detection location, or sample location. Event data can be fertilization, irrigation, plant protection, or harvesting events, and its spatial attribution can be determined based on the scope of action, operational range, or associated range.

[0044] Multi-source data belonging to the same prediction unit and the same observation period are organized into a single data unit group. For example, for prediction unit U1 and observation period T1, remote sensing data covering U1 and collected at time T1, soil moisture data representing U1 and collected at time T1, irrigation event data whose scope includes U1 and whose occurrence time is T1, and sample detection data collected within U1 and whose detection time is T1 can be organized into a data unit group G(U1,T1).

[0045] Through the above methods, data from different collection channels, different spatial forms, and different temporal granularities are converted into unit data groups with clear spatial-temporal affiliations.

[0046] In this embodiment, for the multi-source data in each unit data group, a corresponding data quality identifier is generated based on the quality evaluation information of the multi-source data in the corresponding prediction unit and the corresponding observation period. The data quality identifier is used to characterize the usability of the multi-source data in subsequent prediction processing in the corresponding prediction unit and the corresponding observation period.

[0047] Quality assessment information may include at least two of the following: integrity assessment parameters, timeliness assessment parameters, spatial matching assessment parameters, anomaly assessment parameters, and sample support assessment parameters. Integrity assessment parameters characterize the completeness of multi-source data within the corresponding prediction unit and observation period. Timeliness assessment parameters characterize the temporal matching degree between multi-source data and the corresponding observation period. Spatial matching assessment parameters characterize the spatial matching degree between multi-source data and the corresponding prediction unit. Anomaly assessment parameters characterize the degree of anomaly in the multi-source data. Sample support assessment parameters characterize the degree of support provided by the assessment data or sample data to the corresponding prediction unit and observation period.

[0048] For remote sensing data, integrity evaluation parameters can be determined based on the effective observation ratio, timeliness evaluation parameters can be determined based on the matching degree between image acquisition time and observation period, spatial matching evaluation parameters can be determined based on the overlap between image coverage and prediction unit, and anomaly evaluation parameters can be determined based on cloud occlusion, abnormal pixels, or registration anomalies.

[0049] For ground-collected data, integrity evaluation parameters can be determined based on sampling continuity, timeliness evaluation parameters can be determined based on the degree of matching between sampling time and observation period, spatial matching evaluation parameters can be determined based on the spatial relationship between the collection point location and the prediction unit, and anomaly evaluation parameters can be determined based on sensor anomalies.

[0050] For evaluation data, data quality labels can be generated based on sample size, sample distribution, sample collection time, and sample representativeness. Specifically, sample size and distribution can be used to determine the sample support evaluation parameters, sample collection time can be used to determine the timeliness evaluation parameters, and sample location or coverage area can be used to determine the spatial matching evaluation parameters.

[0051] In some embodiments, the data quality identifier includes a data quality score and a data quality level. The data quality score Q can be determined as follows: Q=wC×C+wT×T+wS×S+wA×A+wN×N; Where C is the integrity evaluation parameter, T is the timeliness evaluation parameter, S is the spatial matching evaluation parameter, A is the anomaly evaluation parameter, and N is the sample support evaluation parameter; wC, wT, wS, wA, and wN are the weight parameters of the corresponding evaluation parameters. The sum of the weight parameters involved in the calculation is 1, and the weight parameters corresponding to the evaluation parameters not involved in the calculation are 0.

[0052] Different types of multi-source data can be evaluated using different combinations of evaluation parameters. For example, remote sensing data can primarily utilize integrity, timeliness, spatial matching, and anomaly evaluation parameters; evaluation data can primarily utilize timeliness, spatial matching, and sample support evaluation parameters. Thus, data quality labels can express the usability of different types of multi-source data in a unified manner within the corresponding prediction unit and observation period.

[0053] In some embodiments, each evaluation parameter can be mapped to the same value range according to a preset normalization rule. The data quality level can be determined based on the comparison between the data quality score and a preset quality threshold, which can be determined based on historical data of the target area, the accuracy requirements of the target task, or the sample calibration results.

[0054] In this embodiment, based on multiple unit data groups corresponding to the same prediction unit, feature extraction and correlation analysis are performed on the state data of the target crop under multiple observation periods to obtain the object stage state of the prediction unit under the corresponding observation period.

[0055] Specifically, multiple data sets corresponding to the same prediction unit are arranged in chronological order according to the observation time period, and the state characteristic parameters of the target crop are extracted from the arranged data sets. These state characteristic parameters may include crop growth parameters, canopy state parameters, late-stage decay parameters, and pre-harvest risk parameters. Crop growth parameters can be obtained from state data or detection data; canopy state parameters can be obtained from remote sensing detection data, UAV detection data, or ground image detection data; late-stage decay parameters and pre-harvest risk parameters can be determined from state data, environmental data, operational data, and historical data.

[0056] Furthermore, based on the changing relationships of state characteristic parameters across multiple observation periods, state change trends are obtained. These trends can characterize whether the target crop's state is improving, stabilizing, declining, or experiencing abnormal fluctuations. The stage state of the target crop is determined by combining environmental assessment parameters derived from environmental data and / or agricultural event parameters derived from operational data. Environmental assessment parameters can be formed from temperature, moisture, light, soil, or meteorological conditions; agricultural event parameters can be formed from events such as sowing, fertilization, irrigation, plant protection, weeding, or harvesting.

[0057] In this embodiment, the target stage state can be the crop growth stage state, such as the seedling stage, vegetative growth stage, flowering stage, grain-filling stage, maturity stage, or pre-harvest stage. The target stage state can be determined based on the state changes of the same prediction unit over multiple observation periods, rather than solely based on a uniform calendar time.

[0058] In some embodiments, stage state mapping rules can be pre-configured. These rules map state change trends, environmental assessment parameters, and agricultural event parameters to the stage state of an object. For example, when state characteristic parameters continuously increase from low values ​​and a sowing event has occurred, it can be identified as the seedling stage or vegetative growth stage; when state characteristic parameters enter a stable or slowly declining phase accompanied by changes in corresponding environmental assessment parameters, it can be identified as the grain-filling stage, maturity stage, or pre-harvest stage.

[0059] In this embodiment, based on the object stage status of the current prediction unit during the current observation period and the data quality identifiers of the multi-source data in the current unit data group, a first task adaptation parameter for the first prediction task and a second task adaptation parameter for the second prediction task are generated respectively.

[0060] Specifically, based on the target stage state, the first-stage matching result of the multi-source data in the current unit data group relative to the first prediction task and the second-stage matching result relative to the second prediction task are determined. The first-stage matching result is used to characterize the degree of matching between the multi-source data and the crop yield prediction task in the current target stage state. The second-stage matching result is used to characterize the degree of matching between the multi-source data and the crop quality prediction task in the current target stage state.

[0061] In crop yield prediction, data on crop growth status, soil moisture, fertilization, canopy establishment, and canopy expansion show a high first-stage matching result for the corresponding crop stage. In crop quality prediction, data on temperature, moisture, late-stage decline, pre-harvest risk, and quality sample data show a high second-stage matching result for the corresponding crop stage.

[0062] Furthermore, the first and second task adaptation parameters are generated by combining data quality identifiers. When a multi-source dataset has a high first-stage matching result with the first prediction task, and its data quality identifier indicates that the data has a high degree of usability in the current prediction unit and current observation period, a higher first task adaptation parameter can be generated. When the multi-source dataset matches the first prediction task, but its data quality identifier indicates a temporal mismatch, spatial mismatch, or a high degree of anomaly, the corresponding first task adaptation parameter can be lowered. The second task adaptation parameter can be generated in a corresponding manner.

[0063] In some embodiments, the first-stage matching result, data quality identifier, and first task association parameter can be mapped to parameter values ​​within a preset range, and the mapped parameter values ​​are weighted and fused to obtain the first task adaptation parameter; similarly, the second-stage matching result, data quality identifier, and second task association parameter can be mapped to parameter values ​​within a preset range, and the mapped parameter values ​​are weighted and fused to obtain the second task adaptation parameter. The first task association parameter and the second task association parameter can be determined by historical data, associated data, or preset matching rules.

[0064] The first and second task adaptation parameters for the same multi-source data can differ. For example, soil moisture data can be used for both crop yield and crop quality prediction, but its effectiveness in both tasks can be determined based on the target stage and data quality indicators. Therefore, the first and second task adaptation parameters can reflect the differentiated requirements of different prediction tasks for the same multi-source data.

[0065] In this embodiment, based on the first task adaptation parameters, the multi-source data in the current unit data group are subjected to matching and fusion processing for the first prediction task to form a first prediction input; based on the second task adaptation parameters, the multi-source data in the current unit data group are subjected to matching and fusion processing for the second prediction task to form a second prediction input.

[0066] Specifically, for the first prediction task, the participation level of multi-source data in the current unit data group in the first prediction input is determined according to the first task adaptation parameters. Multi-source data with higher first task adaptation parameters can have a higher participation level in the first prediction input; multi-source data with lower first task adaptation parameters can be downweighted, suppressed, or excluded from the first prediction input. For the second prediction task, the participation level of multi-source data in the current unit data group in the second prediction input is determined in a corresponding manner according to the second task adaptation parameters.

[0067] In some embodiments, when the data quality identifier of a multi-source data does not meet the preset judgment rules, the participation of that multi-source data in the corresponding prediction input can be reduced, or the multi-source data can be removed from the corresponding prediction input. For example, when a remote sensing detection data does not meet the preset judgment rules due to cloud cover or a large number of abnormal pixels, the participation of that remote sensing detection data in the first prediction input or the second prediction input can be reduced.

[0068] When other multi-source data that meet the conditions exist for the same prediction task, their participation can be increased. For example, when the availability of remote sensing data decreases, the participation of ground image detection data, environmental data, historical state data, or sample detection data with better data quality indicators can be increased for the same prediction task. The amount of increase in the participation of other multi-source data can be determined based on the task adaptation parameters corresponding to the multi-source data whose participation has been reduced or removed, as well as the task adaptation parameters of the other multi-source data relative to the same prediction task.

[0069] In some embodiments, the participation level of multi-source data that has been reduced or removed corresponds to the release of participation level. For other multi-source data that meet the preset judgment rules under the same prediction task, the release of participation level can be allocated according to the task adaptation parameter ratio of the other multi-source data to determine the increase in participation level of each other multi-source data.

[0070] Furthermore, the relationships between the first and second predicted input quantities and the multi-source data and task adaptation parameters used to form the corresponding predicted input quantities are determined. These relationships may include the data sources used in the predicted input quantities, data quality identifiers, the first task adaptation parameter, the second task adaptation parameter, and the degree of participation of the multi-source data in the corresponding predicted input quantities.

[0071] In some embodiments, the association may further include multi-source data identifiers, the prediction unit to which it belongs, the observation period to which it belongs, the object stage status, the prediction task identifier, and the corresponding prediction input identifier, so as to trace and locate the prediction deviation after new evaluation data appears.

[0072] Through the above processing, the first and second prediction inputs are not simply a concatenation of all multi-source data, but rather task-oriented inputs formed based on the object stage status, data quality identifiers, and differences in prediction tasks.

[0073] In this embodiment, the first prediction input and the second prediction input are input into the prediction model to obtain the first prediction result and the second prediction result.

[0074] Specifically, the first prediction input corresponds to the crop yield prediction task, and the prediction model outputs a first prediction result based on the first prediction input. The first prediction result can be yield per unit area, regional yield, yield grade, or yield risk status. The second prediction input corresponds to the crop quality prediction task, and the prediction model outputs a second prediction result based on the second prediction input. The second prediction result can be a quality index value, quality grade, or quality risk status.

[0075] The prediction model can be a model capable of outputting a first prediction result and a second prediction result based on a first prediction input and a second prediction input, respectively. The prediction model can be a prediction model, a parameter generation model, a matching model, or any other model capable of prediction processing. This embodiment does not limit the specific model type, algorithm name, or training method of the prediction model.

[0076] In some embodiments, the prediction model may include processing branches corresponding to the first prediction task and the second prediction task respectively, or it may employ a unified prediction model capable of handling multiple prediction tasks.

[0077] In this embodiment, in response to newly added evaluation data, the newly added evaluation data is compared with the corresponding first prediction result or second prediction result to obtain the prediction deviation; based on the spatial affiliation, temporal affiliation and correlation of the newly added evaluation data, the target prediction unit, target observation period, corresponding object stage state and multi-source data involved in generating the corresponding prediction result corresponding to the prediction deviation are located, and the task adaptation parameters and / or prediction processing parameters under the corresponding prediction task are corrected based on the location results.

[0078] Specifically, when the newly added evaluation data is measured yield data, harvest measurement data, or yield grade evaluation data, the newly added evaluation data is compared with the first prediction result to obtain the first prediction deviation. When the newly added evaluation data is measured quality data, sample test data, or quality grade evaluation data, the newly added evaluation data is compared with the second prediction result to obtain the second prediction deviation.

[0079] The target prediction unit is determined based on the spatial allocation of the newly added evaluation data. For example, the target prediction unit can be determined based on the sample location, yield measurement area, inspection area, or harvest area. The target observation period is determined based on the temporal allocation of the newly added evaluation data. For example, the target observation period can be determined based on the sampling time, inspection time, evaluation time, or harvest time. The corresponding object stage status is determined based on the target prediction unit and the target observation period.

[0080] Based on the correlation between the first or second prediction input and the multi-source data and task adaptation parameters used to form the corresponding prediction input, the multi-source data and their task adaptation parameters involved in generating the corresponding prediction result are located. If the first prediction deviation corresponds to the first prediction result, then the multi-source data and their first task adaptation parameters involved in generating the first prediction result are determined based on the correlation between the first prediction input and the multi-source data and the first task adaptation parameters. If the second prediction deviation corresponds to the second prediction result, then the multi-source data and their second task adaptation parameters involved in generating the second prediction result are determined based on the correlation between the second prediction input and the multi-source data and the second task adaptation parameters.

[0081] During correction, the first task adaptation parameters, second task adaptation parameters, and / or prediction processing parameters can be adjusted based on prediction bias, data quality identifiers, and participation levels. For example, when a first prediction bias indicates a deviation in crop yield prediction, and the bias localization results indicate that a certain multi-source data has a low data quality identifier but a high participation level, the first task adaptation parameters for that multi-source data under the same target stage and the first prediction task can be decreased. When a second prediction bias indicates a deviation in crop quality prediction, and a certain quality-related data has a high data quality identifier but a low participation level, the second task adaptation parameters for that multi-source data under the same target stage and the second prediction task can be increased. If necessary, prediction processing parameters can also be corrected.

[0082] In some embodiments, the prediction processing parameters may include at least one of the following: fusion weights, matching rule parameters, stage matching relationship parameters, basic values ​​of task adaptation parameters, prediction model input parameters, and prediction model parameters, generated during the prediction input formation process. When the prediction bias is mainly concentrated in a single prediction unit, a single observation period, or a single object stage state, the corresponding task adaptation parameters can be corrected first; when the prediction bias shows the same trend in multiple prediction units, multiple observation periods, or multiple object stage states, the prediction processing parameters can be corrected.

[0083] Therefore, the newly added evaluation data is not only used as an overall training sample, but also used to locate the spatial, temporal, object stage state and participating data corresponding to the prediction bias, and to correct the task adaptation parameters and / or prediction processing parameters under the corresponding prediction task.

[0084] In some embodiments, prediction units can be established based on planting boundaries, field boundaries, management zone boundaries, remote sensing identification boundaries, or regions of varying growth. Field units are suitable for scenarios where field boundaries are clear and management measures are relatively uniform; grid units are suitable for scenarios requiring refined spatial prediction; management zone units are suitable for scenarios where irrigation, fertilization, or plant protection strategies differ between zones; and regions of varying growth are suitable for scenarios where crop status differences are significant within the target area. Observation periods can be established based on sowing time, expected harvest time, multi-source data collection frequency, and the time of agricultural events, enabling data from different sources to be aggregated at the same time scale.

[0085] In some embodiments, detection data, status data, environmental data, operational data, historical data, correlation data, and evaluation data may each have different functions. Detection data can be used to obtain detection results for the target crop; status data can be used to characterize changes in crop status; environmental data can be used to characterize the environment in which the crop exists; operational data can be used to characterize agricultural events; historical data can be used to provide historical status, historical yield, or historical quality references; correlation data can be used to determine the matching relationship between data and prediction tasks; and evaluation data can be used to compare with prediction results and trigger correction processes.

[0086] In some embodiments, data with different spatial forms can be assigned using different methods. For areal data such as remote sensing imagery and UAV imagery, spatial assignment can be determined based on the overlap between the coverage area and the prediction unit; for point data such as soil sensor data and sample detection data, spatial assignment can be determined based on the collection location, detection location, or sample location; for event data such as fertilization, irrigation, plant protection, or harvesting, spatial assignment can be determined based on the scope of action, operational range, or associated range. Temporal assignment can be determined based on the collection time, detection time, occurrence time, evaluation time, or associated time.

[0087] In some embodiments, quality assessment information may consist of integrity assessment parameters, timeliness assessment parameters, spatial matching assessment parameters, anomaly assessment parameters, and sample support assessment parameters. These assessment parameters can reflect the usability of multi-source data under the corresponding prediction unit and the corresponding observation period from different perspectives. Integrity assessment parameters reflect whether the data coverage or sampling is sufficient; timeliness assessment parameters reflect whether the data time matches the observation period; spatial matching assessment parameters reflect whether the data location or coverage area matches the prediction unit; anomaly assessment parameters reflect whether there is occlusion, abnormal pixels, sensor anomalies, or recording anomalies in the data; and sample support assessment parameters reflect whether the assessment data or sample data has sufficient quantity and representativeness.

[0088] In some embodiments, the data quality score Q can be generated from multiple evaluation parameters and weighting parameters. Different combinations of evaluation parameters can be used for different types of multi-source data. The data quality score and data quality level can be used to determine the degree to which multi-source data participates in the prediction task. When the data quality score is low or the data quality level does not meet the preset judgment rules, the participation of multi-source data can be reduced.

[0089] In some embodiments, the stage state of an object can be determined by state characteristic parameters of the same prediction unit over multiple observation periods. These state characteristic parameters may include crop growth parameters, canopy state parameters, late-stage decay parameters, and pre-harvest risk parameters. The changing trends of these state characteristic parameters can be combined with environmental assessment parameters and agricultural event parameters to determine the crop growth stage state, thus matching the object stage state with the state changes of the prediction unit itself.

[0090] In some embodiments, the first task adaptation parameters and the second task adaptation parameters can be generated based on the object's stage status and data quality identifier. For crop yield prediction tasks, growth status data, soil moisture data, fertilization data, etc., can have higher first task adaptation parameters; for crop quality prediction tasks, temperature conditions, moisture conditions, late-stage decay status, pre-harvest risk status, quality sample data, etc., can have higher second task adaptation parameters. The same multi-source data can correspond to different first task adaptation parameters and second task adaptation parameters.

[0091] In some embodiments, when the data quality identifier of a certain multi-source data does not meet the preset judgment rules, its participation level can be reduced or it can be removed; when there are other multi-source data that meet the conditions under the same prediction task, the participation level of the other multi-source data can be increased. In this way, the impact of low-quality data on the prediction input can be reduced, and the correlation between the prediction input and the corresponding prediction task can be maintained.

[0092] In some embodiments, when the newly added evaluation data corresponds to a first prediction task, the newly added evaluation data is compared with the first prediction result to obtain a first prediction deviation; when the newly added evaluation data corresponds to a second prediction task, the newly added evaluation data is compared with the second prediction result to obtain a second prediction deviation. The prediction deviation can be combined with the spatial and temporal attribution of the newly added evaluation data, as well as the correlation between the prediction input and multi-source data and task adaptation parameters, to locate the multi-source data involved in generating the corresponding prediction result, and to correct the first task adaptation parameters, the second task adaptation parameters, and / or prediction processing parameters.

[0093] In this embodiment, multi-source data from various sources in agricultural scenarios, including detection, status, environment, operation, history, correlation, and evaluation, can be categorized into prediction units and observation periods to form unit data groups. Data quality identifiers are used to evaluate the usability of multi-source data under current spatial-temporal conditions. The object stage status reflects the state changes of the target crop under the current prediction unit and observation period. First and second task adaptation parameters control the participation of multi-source data in crop yield prediction and crop quality prediction tasks, respectively. New evaluation data is used to locate and correct prediction biases. Therefore, the first and second prediction inputs can be adapted to crop yield prediction and crop quality prediction tasks, respectively, and can be directionally corrected based on measured evaluation results.

[0094] Example 2 In another alternative embodiment, the target area is an industrial production area, such as a production workshop, production line, workstation, process section, or testing area. The target object can be a product batch, a workpiece to be processed, production equipment, or a manufacturing unit. The first prediction task can be product output prediction, capacity prediction, or output risk prediction; the second prediction task can be product quality prediction, defect risk prediction, pass rate prediction, or rework risk prediction.

[0095] The first prediction result can be the product output quantity, production capacity level, or output risk status. The second prediction result can be quality index values, quality level, defect risk status, pass rate, or rework risk status. Multi-source data can include visual inspection data, dimensional inspection data, equipment operation data, process parameters, environmental data, historical output data, historical quality data, and final inspection evaluation data. The object stage status can be the material loading stage, processing stage, assembly stage, inspection stage, rework stage, warehousing stage, stable operation stage, or abnormal fluctuation stage. Newly added evaluation data can be the actual output quantity, final inspection quality results, sampling defect rate, quality level, rework results, or equipment inspection results.

[0096] This embodiment illustrates that in industrial production scenarios with multi-source data, object stage states, multiple prediction tasks, and feedback evaluation data, the same upper-level processing logic as in the aforementioned embodiments can also be used. This embodiment does not constitute a limitation on the application scenarios of this application.

[0097] In this embodiment, multiple prediction units are established according to the spatial evaluation scale of the target area, and multiple observation periods are established according to the periodic state change process of the target object.

[0098] Specifically, prediction units can be established based on the production workshop layout, production line distribution, equipment arrangement, workstation setup, process segment division, or inspection area boundaries. Prediction units can be workstation units, process segment units, inspection area units, equipment units, or manufacturing units. For example, material loading stations, processing stations, assembly stations, inspection stations, rework stations, and warehousing areas can each be used as prediction units; a continuous production line can be divided into multiple prediction units according to process segments; or a manufacturing unit consisting of multiple pieces of equipment can be used as a prediction unit.

[0099] Observation periods can be established based on batch start time, batch end time, equipment sampling frequency, process event occurrence time, inspection time, or shift time. For product batches, observation periods can be established according to the batch flow process; for continuous production lines, observation periods can be established according to hours, shifts, or process cycle time; for process stages, observation periods can be established according to stages such as material loading, processing, assembly, inspection, rework, or warehousing.

[0100] As a result, the industrial production area is divided into multiple prediction units that can respectively carry out data collection, object stage status judgment, prediction processing and feedback correction, and the production process of the target object is divided into multiple observation periods that can be used for time attribution judgment and state change analysis.

[0101] In this embodiment, industrial multi-source data from different acquisition channels are acquired, and the multi-source data are assigned to corresponding prediction units and corresponding observation periods according to their spatial and temporal affiliations, forming multiple unit data groups.

[0102] Inspection data can include visual inspection data, dimensional inspection data, and online inspection data. Status data can include processing status, assembly status, inspection status, and flow status. Environmental data can include temperature, humidity, vibration, cleanliness, and air pressure. Operational data can include equipment operation data, process parameters, downtime records, and production line cycle time data. Historical data can include historical output data, historical quality data, and historical equipment status. Evaluation data can include actual output quantity, final inspection quality results, sampling defect rate, quality grade, and rework results.

[0103] For visual inspection data, spatial attribution can be determined based on the workstation or inspection area where the visual inspection equipment is located, and temporal attribution can be determined based on the image acquisition time or inspection time. For dimensional inspection data, spatial attribution can be determined based on the inspection location, the workpiece's workstation or inspection area, and temporal attribution can be determined based on the inspection time. For equipment operation data, spatial attribution can be determined based on the workstation, process section, or manufacturing unit to which the equipment belongs, and temporal attribution can be determined based on the sampling time. For process parameters, spatial attribution can be determined based on the corresponding equipment, workstation, or process section, and temporal attribution can be determined based on the parameter's effective time or acquisition time. For final inspection and evaluation data, spatial attribution can be determined based on the product batch, inspection area, or quality evaluation source, and temporal attribution can be determined based on the inspection time or evaluation time.

[0104] For example, for prediction unit P1 and observation period T2, the equipment operation data, visual inspection data, dimensional inspection data, process parameters, environmental data, and historical quality data of this process segment within T2 can be organized into a unit data group G(P1,T2). This unit data group is used to characterize the processing status, equipment operation status, environmental status, quality risk status, and historical correlation status of the target object in this process segment and observation period.

[0105] In this embodiment, for the multi-source data in each unit data group, a corresponding data quality identifier is generated based on the quality evaluation information of the multi-source data under the corresponding prediction unit and the corresponding observation period.

[0106] For visual inspection data, data quality labels can be generated based on the effective image ratio, the matching degree between image acquisition time and observation period, the matching relationship between inspection station and prediction unit, and the image anomaly ratio. The effective image ratio can be used to determine integrity evaluation parameters, the image acquisition time can be used to determine timeliness evaluation parameters, the inspection station matching relationship can be used to determine spatial matching evaluation parameters, and the image anomaly ratio can be used to determine anomaly evaluation parameters.

[0107] For equipment operation data, data quality labels can be generated based on data continuity, sampling time matching degree, equipment-related prediction unit matching degree, and sensor abnormal status. For final inspection and evaluation data, data quality labels can be generated based on sample quantity, detection coverage, detection time, and sample representativeness. Among these, sample quantity and sample representativeness can be used to determine the sample-supporting evaluation parameters.

[0108] In some embodiments, data quality identification may include a data quality score and a data quality level. The data quality score can be generated from at least two of the following: integrity evaluation parameters, timeliness evaluation parameters, spatial matching evaluation parameters, anomaly evaluation parameters, and sample support evaluation parameters, along with their weights. Different combinations of evaluation parameters can be used for different industrial multi-source data. For example, visual inspection data may emphasize integrity and anomaly evaluation parameters; equipment operation data may emphasize integrity, timeliness, and anomaly evaluation parameters; and final inspection evaluation data may emphasize spatial matching and sample support evaluation parameters.

[0109] The above method can be used to obtain data quality indicators that characterize the availability of industrial multi-source data in the corresponding prediction unit and the corresponding observation period.

[0110] In some embodiments, evaluation parameters in an industrial scenario can be mapped to the same value range according to a preset normalization rule. The data quality level can be determined based on the comparison between the data quality score and a preset quality threshold, which can be determined based on historical batch data, process segment data distribution, or quality evaluation sample calibration results.

[0111] In this embodiment, based on multiple unit data groups corresponding to the same prediction unit, feature extraction and correlation analysis are performed on the state data of the industrial target object under multiple observation periods to obtain the object stage state of the prediction unit under the corresponding observation period.

[0112] Specifically, multiple data sets corresponding to the same prediction unit are arranged in chronological order according to the observation period, and state characteristic parameters are extracted from the arranged data sets. State characteristic parameters may include processing progress parameters, equipment load parameters, conveying cycle time parameters, inspection pass rate parameters, defect trend parameters, and downtime fluctuation parameters.

[0113] Based on the changing relationship of state characteristic parameters over multiple observation periods, the state change trend can be obtained. For example, when the equipment load parameters remain stable, the conveying cycle parameters fluctuate little, and the inspection pass rate is stable, it corresponds to the stable operation stage; when the downtime fluctuation parameters increase or the defect trend parameters rise, it corresponds to the abnormal fluctuation stage; when the target object is at a specific process flow position, it corresponds to the material loading stage, processing stage, assembly stage, inspection stage, rework stage, or warehousing stage.

[0114] Furthermore, the stage status of the object is determined by combining environmental assessment parameters formed from environmental data and / or operational event parameters formed from operational data. Environmental assessment parameters can be formed from temperature, humidity, vibration, cleanliness, or air pressure; operational event parameters can be formed from material loading events, processing events, assembly events, testing events, rework events, warehousing events, shutdown events, or line changeover events.

[0115] In some embodiments, industrial stage state mapping rules can be pre-configured. These rules map state change trends, environmental assessment parameters, and operational event parameters to object stage states such as stable operation stage, abnormal fluctuation stage, processing stage, testing stage, or rework stage.

[0116] In this embodiment, based on the object stage status of the current prediction unit during the current observation period and the data quality identifiers of the multi-source data in the current unit data group, a first task adaptation parameter for the first prediction task and a second task adaptation parameter for the second prediction task are generated respectively.

[0117] When the first forecasting task is product output forecasting, capacity forecasting, or output risk forecasting, the first-stage matching results relative to the first forecasting task can be determined based on the target stage status, including equipment operation data, production line cycle time data, downtime records, process parameters, and historical output data. When the second forecasting task is product quality forecasting, defect risk forecasting, pass rate forecasting, or rework risk forecasting, the second-stage matching results relative to the second forecasting task can be determined based on the target stage status, including visual inspection data, dimensional inspection data, equipment vibration data, environmental data, historical quality data, and final inspection evaluation data.

[0118] For example, during the processing stage, equipment operation data, production line cycle time data, and process parameters are typically strongly correlated with product output prediction tasks; during the inspection stage or abnormal fluctuation stage, visual inspection data, dimensional inspection data, equipment vibration data, and historical quality data are typically strongly correlated with product quality prediction tasks. The system further combines data quality identifiers to generate first task adaptation parameters and second task adaptation parameters.

[0119] In some embodiments, the first-stage matching results, data quality identifiers, and first task-related parameters can be parameter-mapped and weighted to obtain first task adaptation parameters; the second-stage matching results, data quality identifiers, and second task-related parameters can be parameter-mapped and weighted to obtain second task adaptation parameters. The first and second task-related parameters can be determined from historical output data, historical quality data, process-related data, or preset matching rules.

[0120] The first task adaptation parameters and the second task adaptation parameters corresponding to the same multi-source data can be different. For example, equipment operation data can affect both product output forecasting and product quality forecasting, but its first task adaptation parameter in the product output forecasting task can be different from its second task adaptation parameter in the product quality forecasting task.

[0121] In this embodiment, a first prediction input quantity is formed based on the first task adaptation parameters for product output prediction, capacity prediction, or output risk prediction, and a second prediction input quantity is formed based on the second task adaptation parameters for product quality prediction, defect risk prediction, pass rate prediction, or rework risk prediction.

[0122] Specifically, the first predictive input can be formed from data that matches the first predictive task and whose data quality indicators meet the requirements, including equipment operation data, production line cycle time data, downtime records, process parameters, historical output data, and environmental data. The second predictive input can be formed from data that matches the second predictive task and whose data quality indicators meet the requirements, including visual inspection data, dimensional inspection data, equipment vibration data, environmental data, historical quality data, and final inspection evaluation data.

[0123] When visual inspection data is abnormal, equipment operation data sampling is interrupted, dimensional inspection samples are insufficient, final inspection evaluation data coverage is inadequate, or process parameter records are abnormal, the participation of that multi-source data in the corresponding predicted input can be reduced or eliminated based on data quality indicators. When other multi-source data that meet the conditions exist under the same prediction task, the participation of other multi-source data in the corresponding predicted input can be increased. For example, when a certain visual inspection data is abnormal, the participation of dimensional inspection data, historical quality data, or final inspection evaluation data under the same quality prediction task can be increased; when the sampling of a certain equipment operation data is interrupted, the participation of production line cycle time data, downtime records, or historical output data under the same output prediction task can be increased.

[0124] In some embodiments, the participation level of industrial multi-source data that has been reduced or removed is corresponding to the release of participation level. For other industrial multi-source data that meet the preset judgment rules under the same prediction task, the release of participation level can be allocated according to the task adaptation parameter ratio of the other industrial multi-source data to determine the amount of increase in participation level for each other industrial multi-source data.

[0125] Furthermore, the relationships between the first and second predictive inputs and the multi-source data and task adaptation parameters used to form the corresponding predictive inputs are determined. These relationships can be used to characterize the source, data quality identifier, task adaptation parameters, and participation level of each multi-source data in the predictive inputs.

[0126] In this embodiment, the first prediction input and the second prediction input are input into the prediction model to obtain the first prediction result and the second prediction result.

[0127] Specifically, the first prediction result can be the product output quantity, production capacity level, or output risk status. The second prediction result can be a quality indicator value, quality level, defect risk status, pass rate, or rework risk status. The prediction model can be a model capable of receiving the first and second prediction inputs and outputting the first and second prediction results respectively. This embodiment does not limit the specific prediction model type, specific algorithm name, or specific equipment model.

[0128] In some embodiments, the prediction model may include processing branches corresponding to the first prediction task and the second prediction task respectively, or it may employ a unified prediction model capable of handling multiple prediction tasks.

[0129] In this embodiment, in response to newly added evaluation data, the newly added evaluation data is compared with the corresponding first prediction result or second prediction result to obtain the prediction deviation; based on the spatial affiliation, temporal affiliation and correlation of the newly added evaluation data, the target prediction unit, target observation period, corresponding object stage state and multi-source data involved in generating the corresponding prediction result corresponding to the prediction deviation are located, and the task adaptation parameters and / or prediction processing parameters under the corresponding prediction task are corrected based on the location results.

[0130] When the newly added evaluation data is the actual output quantity, capacity statistics, or output risk assessment result, the newly added evaluation data is compared with the first prediction result to obtain the first prediction deviation. When the newly added evaluation data is the final inspection quality result, sampling defect rate, quality grade, rework result, or equipment inspection result, the newly added evaluation data is compared with the second prediction result to obtain the second prediction deviation.

[0131] The target prediction unit is determined based on the production line, workstation, process section, inspection area, or manufacturing unit corresponding to the newly added evaluation data. The target observation period is determined based on batch time, inspection time, shift time, or process time. The corresponding object stage status is determined based on the target prediction unit and the target observation period.

[0132] Based on the correlation between the first or second prediction input and the multi-source data and task adaptation parameters used to form the corresponding prediction input, the multi-source data involved in generating the corresponding prediction result is located. If the first prediction deviation corresponds to a product output prediction task, the first task adaptation parameters of the equipment operation data, production line cycle time data, downtime records, or process parameters involved in generating the first prediction result can be adjusted in the target prediction unit, target observation period, and corresponding object stage state. If the second prediction deviation corresponds to a product quality prediction task, the second task adaptation parameters of the visual inspection data, dimensional inspection data, equipment vibration data, environmental data, or historical quality data involved in generating the second prediction result can be adjusted in the target prediction unit, target observation period, and corresponding object stage state. If necessary, the prediction processing parameters can be corrected based on the prediction deviation, data quality identifier, and participation level.

[0133] In some embodiments, the prediction processing parameters may include at least one of the following: fusion weights, matching rule parameters, stage matching relationship parameters, basic values ​​of task adaptation parameters, prediction model input parameters, and prediction model parameters, during the formation of the prediction input. When the prediction deviation can be located to specific industrial multi-source data, the task adaptation parameters corresponding to that industrial multi-source data can be corrected first; when the prediction deviation shows a systematic trend across multiple workstations, multiple observation periods, or multiple object stage states, the prediction processing parameters can be corrected.

[0134] Therefore, the newly added evaluation data, such as the actual output quantity, final inspection quality results, sampling defect rate, or rework results, can be used to locate the spatial, temporal, stage status, and participating data corresponding to the prediction deviation, and to correct the task adaptation parameters and / or prediction processing parameters under the corresponding prediction task.

[0135] In some embodiments, production lines, workstations, process sections, inspection areas, or equipment units can form prediction units; batch start time, batch end time, equipment sampling frequency, process event occurrence time, inspection time, or shift time can form observation periods. Thus, industrial production data can be organized into a unified space-time processing framework.

[0136] In some embodiments, visual inspection data, dimensional inspection data, and online inspection data can be used as inspection data; processing status, assembly status, inspection status, and flow status can be used as status data; temperature, humidity, vibration, cleanliness, and air pressure status can be used as environmental data; equipment operation data, process parameters, downtime records, and production line cycle time data can be used as operation data; historical output data, historical quality data, and historical equipment status can be used as historical data; and actual output quantity, final inspection quality results, sampling defect rate, quality grade, and rework results can be used as evaluation data.

[0137] In some embodiments, detection area data or process section coverage data can be treated as area data, and spatial attribution is determined based on the overlap between the coverage area and the prediction unit; workstation detection data, equipment sensor data, or dimensional detection point data can be treated as point data, and spatial attribution is determined based on the acquisition location, detection location, or equipment location; line changeover events, downtime events, processing events, rework events, or detection events can be treated as event data, and spatial attribution is determined based on the event's scope of action, operating range, or associated range. Temporal attribution can be determined based on the acquisition time, detection time, occurrence time, evaluation time, or associated time.

[0138] In some embodiments, integrity evaluation parameters can evaluate whether the visual inspection data has sufficiently valid images, whether the equipment operation data has sufficient sampling continuity, and whether the final inspection evaluation data has a sufficient number of samples. Timeliness evaluation parameters can evaluate the matching relationship between image acquisition time, equipment sampling time, or final inspection time and the observation period. Spatial matching evaluation parameters can evaluate the matching relationship between the inspection station, the area to which the equipment belongs, or the inspection area and the prediction unit. Anomaly evaluation parameters can evaluate image anomalies, sensor anomalies, equipment anomalies, or process parameter anomalies. Sample support evaluation parameters can evaluate the number of final inspection samples, the inspection coverage, and the representativeness of the samples.

[0139] In some embodiments, the data quality score Q can be generated from the integrity evaluation parameter C, the timeliness evaluation parameter T, the spatial matching evaluation parameter S, the anomaly evaluation parameter A, and the sample support evaluation parameter N, along with their weight parameters. Visual inspection data can focus on the integrity and anomaly evaluation parameters; equipment operation data can focus on the integrity, timeliness, and anomaly evaluation parameters; and final inspection evaluation data can focus on the spatial matching and sample support evaluation parameters.

[0140] In some embodiments, the stage status of an object can be determined based on state characteristic parameters of the same prediction unit over multiple observation periods. State characteristic parameters may include processing progress parameters, equipment load parameters, conveying cycle time parameters, inspection pass rate parameters, defect trend parameters, and downtime fluctuation parameters. The changing trends of the state characteristic parameters can be combined with environmental assessment parameters and operational event parameters to determine the stable operation stage, abnormal fluctuation stage, processing stage, inspection stage, or rework stage.

[0141] In some embodiments, the first task adaptation parameters and the second task adaptation parameters can be generated based on the object stage status and data quality identifiers. For product output prediction tasks, equipment operation data, production line cycle time data, downtime records, process parameters, and historical output data can have higher first task adaptation parameters. For product quality prediction tasks, visual inspection data, dimensional inspection data, equipment vibration data, environmental data, historical quality data, and final inspection evaluation data can have higher second task adaptation parameters. The same equipment operation data or environmental data can correspond to different first task adaptation parameters and second task adaptation parameters.

[0142] In some embodiments, when visual inspection data is abnormal, equipment operation data sampling is interrupted, dimensional inspection samples are insufficient, final inspection evaluation data coverage is inadequate, or process parameter records are abnormal, the participation of such multi-source data in the corresponding predicted input can be reduced, or it can be removed from the corresponding predicted input. When other multi-source data that meet the conditions exist under the same prediction task, the participation of other multi-source data in the corresponding predicted input can be increased to reduce the impact of low-quality industrial data on the predicted input.

[0143] In some embodiments, when the newly added evaluation data is the actual output quantity, capacity statistics, or output risk assessment result, it can be compared with the first prediction result to obtain the first prediction deviation; when the newly added evaluation data is the final inspection quality result, sampling defect rate, quality grade, rework result, or equipment inspection result, it can be compared with the second prediction result to obtain the second prediction deviation. Based on the spatial and temporal attribution of the newly added evaluation data and the correlation between the prediction input and multi-source data and task adaptation parameters, the multi-source data involved in generating the corresponding prediction result can be located, and the first task adaptation parameters, the second task adaptation parameters, and / or prediction processing parameters can be corrected.

[0144] Through this embodiment, visual inspection data, dimensional inspection data, equipment operation data, process parameters, environmental data, historical output data, historical quality data, and final inspection evaluation data in industrial production scenarios can be formed into unit data groups according to prediction units and observation periods. The first prediction input quantity corresponding to the product output prediction task and the second prediction input quantity corresponding to the product quality prediction task can be formed respectively through data quality identification, object stage status, and task adaptation parameters.

[0145] Furthermore, once new evaluation data such as actual output quantity, final inspection quality results, sampling defect rate, quality grade, or rework results are generated, the system can locate the target prediction unit, target observation period, corresponding object stage status, and multi-source data involved in generating the corresponding prediction results based on spatial attribution, temporal attribution, and correlation, and perform correction processing accordingly. Thus, this embodiment demonstrates that the higher-level processing logic of claim 1 is not only applicable to agricultural scenarios but also to target object prediction scenarios in industrial production with multi-source data, stage status, multiple prediction tasks, and feedback evaluation data.

[0146] The above embodiments are only used to illustrate optional implementations of this application in different application scenarios and do not constitute a limitation on the scope of protection of this application. Without departing from the technical concept of this application, those skilled in the art can make adaptive adjustments to the specific forms of the target area, target object, multi-source data type, object stage status, prediction task, and newly added evaluation data according to the specific scenario.

[0147] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0148] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A multi-task prediction method for target objects, characterized in that, include: S100 establishes multiple prediction units according to the spatial evaluation scale of the target area, and establishes multiple observation periods according to the periodic state change process of the target object; S200: Acquire multi-source data from different acquisition channels, and classify the multi-source data into corresponding prediction units and corresponding observation periods according to the spatial and temporal attribution of the multi-source data to form multiple unit data groups; wherein, each unit data group corresponds to a prediction unit and an observation period, and the multi-source data includes at least two types of data used to characterize the state of the target object, the environment in which the target object is located, the operation process of the target object, and the evaluation results of the target object; S300, for the multi-source data in each unit data group, a corresponding data quality identifier is generated based on the quality evaluation information of the multi-source data in the corresponding prediction unit and the corresponding observation period; wherein, the data quality identifier is used to characterize the availability of the multi-source data in the corresponding prediction unit and the corresponding observation period for participating in subsequent prediction processing; S400, based on multiple unit data groups corresponding to the same prediction unit, feature extraction and correlation analysis are performed on the state data of the target object under multiple observation periods to obtain the object stage state of the prediction unit under the corresponding observation period. S500, based on the object stage status of the current prediction unit in the current observation period and the data quality identifier of the multi-source data in the current unit data group, generate a first task adaptation parameter for the first prediction task and a second task adaptation parameter for the second prediction task respectively; wherein, the first task adaptation parameter is used to characterize the effectiveness of the corresponding multi-source data participating in the first prediction task, and the second task adaptation parameter is used to characterize the effectiveness of the corresponding multi-source data participating in the second prediction task. S600: Based on the first task adaptation parameters, perform matching and fusion processing on the multi-source data in the current unit data group for the first prediction task to form a first prediction input; based on the second task adaptation parameters, perform matching and fusion processing on the multi-source data in the current unit data group for the second prediction task to form a second prediction input; and determine the correlation between the first prediction input and the second prediction input and the multi-source data and task adaptation parameters used to form the corresponding prediction inputs. S700, input the first prediction input and the second prediction input into the prediction model to obtain the first prediction result and the second prediction result; S800, in response to newly added evaluation data, compare the newly added evaluation data with the corresponding first prediction result or second prediction result to obtain the prediction deviation; based on the spatial and temporal attribution of the newly added evaluation data and the correlation relationship, locate the target prediction unit, target observation period, corresponding object stage state and multi-source data involved in generating the corresponding prediction result corresponding to the prediction deviation, and perform correction processing on the task adaptation parameters and / or prediction processing parameters under the corresponding prediction task based on the location results.

2. The method according to claim 1, characterized in that, In step S100, multiple prediction units are established according to the spatial evaluation scale of the target area, including: Obtain the spatial boundary information and spatial evaluation requirements of the target area; Based on the spatial boundary information and the spatial evaluation requirements, the target area is divided into multiple regional units that can respectively perform data collection, state judgment, prediction processing and feedback correction, and the regional units are used as the prediction units. In step S100, multiple observation periods are established according to the periodic state change process of the target object, including: The data processing interval is determined based on at least one of the following: the cycle start time, cycle end time, multi-source data acquisition frequency, and runtime event occurrence time of the target object. The multiple observation periods are established according to the data processing intervals.

3. The method according to claim 1, characterized in that, In step S200, the multi-source data includes at least two of the following: detection data, status data, environmental data, operational data, historical data, correlation data, and evaluation data. The detection data is used to characterize the detection results of the target object; The state data is used to characterize the state changes of the target object under the corresponding prediction unit and the corresponding observation period; The environmental data is used to characterize the environmental state of the environment in which the target object is located; The operational data is used to characterize operational events or processing events that act on the target object; The historical data is used to characterize the state changes or prediction results of the target object during the historical observation period; The associated data is used to characterize the relationship between the target object and the prediction task; The evaluation data is used as a basis for feedback correction of the first or second prediction result.

4. The method according to claim 1, characterized in that, In step S200, based on the spatial and temporal attribution of the multi-source data, the multi-source data is assigned to the corresponding prediction unit and the corresponding observation period, including: When the multi-source data is planar data, the corresponding prediction unit is determined according to the overlap relationship between the coverage area of ​​the planar data and the prediction unit. When the multi-source data is point data, the corresponding prediction unit is determined according to the positional relationship between the acquisition location, detection location or sample location of the point data and the prediction unit; When the multi-source data is event data, the corresponding prediction unit is determined according to the scope of action, scope of operation, or scope of association of the event data. The corresponding observation period is determined based on the acquisition time, detection time, occurrence time, evaluation time, or correlation time of the multi-source data.

5. The method according to claim 1, characterized in that, In step S300, the quality evaluation information includes at least two of the following: integrity evaluation parameters, timeliness evaluation parameters, spatial matching evaluation parameters, anomaly evaluation parameters, and sample support evaluation parameters. The integrity evaluation parameter is used to characterize the data integrity of the multi-source data in the corresponding prediction unit and the corresponding observation period. The timeliness evaluation parameter is used to characterize the degree of time matching between the multi-source data and the corresponding observation period; The spatial matching evaluation parameter is used to characterize the degree of spatial matching between the multi-source data and the corresponding prediction unit; The anomaly evaluation parameters are used to characterize the degree of anomaly in the multi-source data; The sample support evaluation parameters are used to characterize the degree of support that the evaluation data or sample data provides for the corresponding prediction unit and the corresponding observation period. The data quality identifier is generated based on at least two of the quality evaluation information items and their corresponding weight parameters.

6. The method according to claim 5, characterized in that, In step S300, the data quality identifier includes a data quality score and a data quality level determined based on the data quality score; The data quality score Q is determined in the following manner: Q=wC×C+wT×T+wS×S+wA×A+wN×N; Wherein, C is the integrity evaluation parameter, T is the timeliness evaluation parameter, S is the spatial matching evaluation parameter, A is the anomaly evaluation parameter, and N is the sample support evaluation parameter; wC, wT, wS, wA, and wN are the weight parameters for the corresponding evaluation parameters; The sum of the weight parameters involved in the calculation is 1, and the weight parameters corresponding to the evaluation parameters not involved in the calculation are 0.

7. The method according to claim 1, characterized in that, In step S400, based on multiple unit data groups corresponding to the same prediction unit, feature extraction and correlation analysis are performed on the state data of the target object under multiple observation periods to obtain the object stage state of the prediction unit under the corresponding observation period, including: Multiple data sets corresponding to the same prediction unit are arranged in chronological order of the observation periods; Extract the state feature parameters of the target object from the arranged multiple cell data groups; Based on the changing relationship of the state characteristic parameters over multiple observation periods, the state change trend is obtained; Based on the state change trend, and in conjunction with environmental assessment parameters formed from environmental data and / or operational event parameters formed from operational data, the object stage state of the prediction unit under the corresponding observation period is determined.

8. The method according to claim 1, characterized in that, In step S500, first task adaptation parameters for the first prediction task and second task adaptation parameters for the second prediction task are generated, including: Based on the object stage state, determine the first stage matching result of the multi-source data in the current unit data group relative to the first prediction task, and the second stage matching result relative to the second prediction task. Based on the data quality identifier, determine the quality evaluation result of the multi-source data in the current prediction unit and the current observation period; Based on the first-stage matching results and the quality evaluation results, the first task adaptation parameters are generated. Based on the matching results of the second stage and the quality evaluation results, the second task adaptation parameters are generated; Among them, the first task adaptation parameters and the second task adaptation parameters corresponding to the same multi-source data can be different.

9. The method according to claim 1, characterized in that, In step S600, the multi-source data in the current unit data group undergoes matching and fusion processing oriented towards the first prediction task or the second prediction task, including: When the data quality identifier of any multi-source data does not meet the preset judgment rules, reduce the participation of that multi-source data in the corresponding prediction input, or remove that multi-source data from the corresponding prediction input. When there are other multi-source data in the current unit data group that are related to the same prediction task and whose data quality identifiers meet the preset judgment rules, the participation of the other multi-source data in the corresponding prediction input is increased. The increase in the participation level of the other multi-source data is determined based on the task adaptation parameters of the multi-source data whose participation level has been reduced or which has been removed, as well as the task adaptation parameters of the other multi-source data relative to the same prediction task.

10. The method according to claim 1, characterized in that, In step S800, the task adaptation parameters and / or prediction processing parameters for the corresponding prediction task are corrected based on the positioning results, including: When the newly added evaluation data corresponds to the first prediction task, the newly added evaluation data is compared with the first prediction result to obtain the first prediction deviation, and the first task adaptation parameters of the multi-source data that participated in generating the first prediction result are adjusted according to the first prediction deviation in the target prediction unit, the target observation period and the corresponding object stage state. When the newly added evaluation data corresponds to the second prediction task, the newly added evaluation data is compared with the second prediction result to obtain the second prediction deviation, and the second task adaptation parameters of the multi-source data involved in generating the second prediction result are adjusted according to the second prediction deviation in the target prediction unit, the target observation period and the corresponding object stage state. The adjustment amount of the first task adaptation parameter or the second task adaptation parameter is determined based on the corresponding prediction deviation, the data quality identifier of the corresponding multi-source data, and the degree of participation of the corresponding multi-source data in the prediction input. In some embodiments, the prediction processing parameters include at least one of the following: fusion weights in the prediction input formation process, matching rule parameters, stage matching relationship parameters, basic values ​​of task adaptation parameters, prediction model input parameters, and prediction model parameters.

Images