Mine dump truck battery life prediction method and system based on operation data

Abstract

The present application relates to the technical field of machine learning, in particular to a mine dump truck battery life prediction method and system based on operation data, comprising the following steps: obtaining battery state data and vehicle operation data and generating an operation data set according to the time stamp, dividing the mine transportation operation segment according to the load change, current direction, state of charge change and vehicle driving state, extracting the periodic aging stress characteristics from the mine transportation operation segment, generating the sample reliability weight according to the consistency of the battery state data and the vehicle operation data, inputting the periodic aging stress characteristics and the sample reliability weight into the prediction model containing the LSTM hidden layer and outputting the prediction result, updating the correction coefficient or input weight according to the difference between the subsequent operation data and the prediction result. The present application forms a corresponding relationship between the life prediction input and the working condition, data quality and feedback through working condition segment division, reliability weighting and residual correction.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and system for predicting the battery life of mining dump trucks based on operational data. Background Technology

[0002] In the field of machine learning technology, data acquisition, labeling, storage, computation, and result output are carried out around sample data, feature variables, training programs, and inference programs. Terminal devices, servers, databases, and sensors transmit operation records through interfaces, buses, or networks. The processing program establishes the computational relationship between input and output based on numerical fields, time series relationships, and label information in historical samples. Among them, the traditional method and system for predicting the battery life of mining dump trucks refers to the technology of predicting the usage status and life records of power batteries under mining transportation conditions. The battery management system collects operation data such as voltage, current, temperature, state of charge, number of charge and discharge cycles, mileage, load, and working time. This data is written to the local storage area or uploaded to the server via the vehicle controller or on-board terminal. The prediction program calls historical operation records and battery detection records, performs feature processing, model training, and inference calculations according to time series, operating condition category, charge and discharge interval, and life label, and outputs the battery life status or estimated remaining usage time.

[0003] Traditional battery life prediction processes for mining dump trucks use records of voltage, current, temperature, state of charge, number of charge / discharge cycles, mileage, load, and operating time as model inputs. The correspondence between continuous records and operational processes such as loading, heavy-load driving, empty return, charging, and parking is unclear. The consistency difference between battery operating status and vehicle operating status is difficult to use in judging sample quality. The prediction results lack correction basis based on the source of operating conditions, making it difficult to distinguish the data impact of different transportation conditions in life state estimation and remaining service time estimation. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting the battery life of mining dump trucks based on operational data. This method establishes a correspondence between the battery status data of mining dump trucks and vehicle operation data according to the mining transportation operation process. It also incorporates aging stress, data consistency, and subsequent operational feedback in different operation segments into the life prediction process, providing a technical solution that adapts to the characteristics of operational data for predicting the battery life status or remaining service time under mining transportation conditions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution, providing a method for predicting the battery life of a mining dump truck based on operational data, comprising the following steps: acquiring battery status data and vehicle operation data output by the vehicle controller or on-board terminal, and generating an operational data set according to timestamps; dividing the operational data set into multiple mining transportation operation segments based on load changes, current direction, state of charge changes, and vehicle driving status; and extracting current continuity characteristics, temperature range characteristics, state of charge swing characteristics, load continuity characteristics, energy consumption per unit mileage characteristics, depth of charge characteristics, and resting voltage recovery characteristics from the mining transportation operation segments. Complex features are used to generate periodic aging stress features. When the battery state data and the vehicle operation data are consistent in time, direction, and state, the sample confidence weight is increased; when any consistency condition is not met, the sample confidence weight is decreased. The periodic aging stress features and the sample confidence weight are input into a prediction model containing an LSTM hidden layer, and the battery life state or remaining usage time prediction result is output. The prediction residual is generated based on the difference between the subsequent operation data or battery detection records and the prediction result, and the correction coefficient or input weight in the subsequent prediction is updated according to the operation segment type corresponding to the prediction residual.

[0006] Furthermore, the battery status data includes voltage, current, temperature, state of charge, and number of charge / discharge cycles, and the vehicle operation data includes mileage, load, operation time, and vehicle driving status. When generating the operation data set according to timestamps, data rows are established using the battery status data and the vehicle operation data within the same collection period. When the collection time difference between the battery status data and the vehicle operation data meets the preset synchronization interval, the data row is retained; when the collection time difference does not meet the preset synchronization interval, the corresponding data row is discarded or marked as a data row to be verified.

[0007] Preferably, when dividing the operational data set into multiple mining transportation operation segments, the retained data rows and the data rows to be verified are read in chronological order; when the load change between adjacent data rows meets the loading or unloading state change conditions, the segment boundary is determined; when the current direction between adjacent data rows and the charge state change trend meet the charging or discharging direction conditions, the charging segment or the operating segment is determined; when the vehicle driving state meets the parking conditions and the current is in the static range, the parking segment is determined; based on the segment boundary, the charging segment, the operating segment, and the parking segment, the segment type, start time, and end time of the mining transportation operation segment are generated.

[0008] Furthermore, when generating the periodic aging stress characteristics, the current duration range, temperature duration range, state of charge variation range, load duration, mileage variation, and charging depth are statistically analyzed according to the segment type of the mining transportation operation segment. When the mining transportation operation segment is a running segment, the energy consumption variation per unit mileage and load duration are written into the segment-level aging stress characteristics. When the mining transportation operation segment is a charging segment, the charging depth and temperature duration range are written into the segment-level aging stress characteristics. When the mining transportation operation segment is a parking segment, the static voltage recovery characteristics are written into the segment-level aging stress characteristics. The periodic aging stress characteristics are generated according to the segment type and time sequence within the same operation cycle based on the segment-level aging stress characteristics.

[0009] Furthermore, when generating the sample credibility weight, the battery status data and the vehicle operation data are considered to be time-consistent when their timestamps are continuous and the acquisition time difference meets a preset synchronization interval; the direction of charge state change is considered to be direction-consistent when it matches the charging and discharging direction corresponding to the current direction; the state is considered to be consistent when the load change, vehicle driving state, and current duration interval meet a preset working condition matching relationship; a first sample credibility weight is generated when time consistency, direction consistency, and state consistency are all met; a second sample credibility weight lower than the first sample credibility weight is generated when any one of the determinations is not met, and the second sample credibility weight is written into the corresponding mine transportation operation segment.

[0010] Furthermore, when the periodic aging stress features and the sample confidence weights are input into the prediction model containing the LSTM hidden layer, the periodic aging stress features within the same operating cycle are formed into a feature sequence according to the segment type, start time, and end time; participation weights are set for the segment-level features in the feature sequence according to the sample confidence weights corresponding to the feature sequence; the feature sequence with the participation weights is input into the prediction model containing the LSTM hidden layer to obtain at least one of the battery life state and remaining usage time as the prediction result.

[0011] Furthermore, when generating the prediction residual, a correspondence is established between the prediction output time, operation cycle, and operation segment type corresponding to the prediction result and subsequent operation data or battery detection records; when the recording time of the subsequent operation data or the battery detection record falls within the subsequent correction cycle corresponding to the prediction result, the difference between the prediction result and the life status in the subsequent operation data or battery detection record is calculated; the correction coefficient or input weight in the subsequent prediction is updated according to the operation segment type corresponding to the difference.

[0012] Furthermore, when updating the correction coefficient or the input weight, the residual records of historical job segments of the same type are read according to the job segment type corresponding to the predicted residual; when the residual records of historical job segments of the same type are consistent with the offset direction of the predicted residual, the correction coefficient of the corresponding job segment type in subsequent prediction is increased; when the residual records of historical job segments of the same type are inconsistent with the offset direction of the predicted residual, the input weight of the corresponding job segment type in subsequent prediction is decreased or the input weight is kept unchanged.

[0013] Optionally, when a data row to be verified has missing or abrupt field changes, the most recent data row of the same mining dump truck that meets the preset synchronization interval is read as the first alternative data source; when the first alternative data source does not have a data row of the same fragment type as the data row to be verified, the data row of the same fragment type in the historical operation fragment of the same type of mining dump truck is read as the second alternative data source; an alternative input is generated according to the first alternative data source or the second alternative data source, and the sample confidence weight of the alternative input is set to be lower than the sample confidence weight corresponding to the data row that meets the preset synchronization interval.

[0014] This invention also provides a battery life prediction system for mining dump trucks based on operational data. This system supports the aforementioned method for predicting the battery life of mining dump trucks. The system includes an operational data access unit, a working condition segment generation unit, an aging stress feature generation unit, a sample credibility processing unit, a life prediction processing unit, and a prediction feedback correction unit. The operational data access unit receives battery status data and vehicle operation data and writes them to a unified storage area according to timestamps. The working condition segment generation unit generates mining transportation operation segments based on load changes, current direction, state of charge changes, and vehicle driving status. The aging stress feature generation unit extracts segment features from the mining transportation operation segments and forms periodic aging stress features. The sample credibility processing unit generates sample credibility weights based on the time consistency, direction consistency, and state consistency between the battery status data and the vehicle operation data. The life prediction processing unit inputs the periodic aging stress features and the sample credibility weights into a prediction model containing an LSTM hidden layer and outputs a battery life status or remaining usage time prediction result. The prediction feedback correction unit updates the correction coefficients or input weights based on the differences between subsequent operational data or battery detection records and the prediction results.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, an operational data set is generated by timestamp, and mining transportation operation segments are divided according to load changes, current direction, state of charge changes, and vehicle driving status. Battery status data corresponds to the vehicle operation process, reducing the confusion of working conditions caused by direct input of a single continuous record. By extracting periodic aging stress features, a prediction model containing LSTM hidden layers obtains inputs containing working condition sources. The credibility weights of generated samples are determined by time consistency, direction consistency, and state consistency, and the training and inference processes can distinguish the degree of data participation in different segments. The correction coefficients or input weights are updated by predicting residuals, and subsequent prediction results obtain operational feedback constraints. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is the main flowchart of the prediction process for the battery life of a mining dump truck according to the present invention; Figure 2 This is a flowchart illustrating the data alignment and abnormal data rollback process for this invention. Figure 3 This is a logic diagram for generating mining transportation operation segments in this invention; Figure 4 A data flow diagram was generated to represent the aging stress characteristics and confidence weights of this invention. Figure 5 This is the closed-loop diagram of the predicted residual feedback correction of the present invention; Figure 6 This is a diagram showing the cooperative relationship between the units of the prediction system of this invention. Detailed Implementation

[0018] The following description, in conjunction with the scope of protection text and the invention content, illustrates the technical solution described in this invention. In subsequent embodiments, battery state data includes voltage, current, temperature, state of charge, and number of charge / discharge cycles; vehicle operation data includes mileage, load, operation duration, and vehicle driving status; the data set formed according to timestamps is uniformly referred to as the operational data set; the segments obtained from the mining transportation operation process are uniformly referred to as mining transportation operation segments; the features obtained by aggregating segment-level aging stress characteristics within the operation cycle are uniformly referred to as cycle-level aging stress characteristics; the weights used to characterize the degree of data participation are uniformly referred to as sample confidence weights; the model output is uniformly referred to as the prediction result; and the feedback quantity formed based on the difference between subsequent records and previous outputs is uniformly referred to as the prediction residual.

[0019] Example 1 Please see Figures 1 to 5 This embodiment provides a method for predicting the battery life of mining dump trucks based on operational data, applicable to the prediction and processing of the life status or remaining usage time of power batteries under mining transportation conditions. The method uses battery status data and vehicle operation data as input sources. After timestamp alignment, an operational data set is formed. Then, based on load changes, current direction, state of charge changes, and vehicle driving status, continuous records are converted into mining transportation operation segments. Operational stress features related to battery aging are extracted from different segments. These periodic aging stress features and sample confidence weights are then input into a prediction model containing an LSTM hidden layer, outputting the prediction result. Subsequent collected operational records or detection records, after differing from the previous output, are used to update the correction coefficients or input weights in subsequent predictions, thus forming a complete processing chain from data acquisition, segmentation, feature generation, confidence weighting, model inference to feedback correction. In this processing chain, machine learning is not executed independently of operational data but is embedded between operational condition segmentation, feature sequence generation, weight binding, and residual feedback.

[0020] S1 acquires battery status data and vehicle operation data output by the vehicle controller or on-board terminal, and generates an operational data set according to timestamps. S1 provides the raw field foundation for the machine learning input preparation stage. Battery status data comes from the voltage, current, temperature, state of charge, and number of charge / discharge cycles recorded during the operation of the mining dump truck's power battery. Vehicle operation data comes from the mileage, load, operation time, and vehicle driving status recorded by the vehicle controller or on-board terminal. When executing S1, the same vehicle, the same battery object, and the same acquisition period are used as data aggregation boundaries. Fields are read in time-stamp order to form data rows that can be traced in chronological order. Each data row contains the voltage, current, temperature, state of charge, number of charge / discharge cycles, mileage, load, operation time, and vehicle driving status within the corresponding acquisition period. The data rows are written into the operational data set for subsequent segmentation and retrieval. The aggregated data rows also serve as the basic records for constructing machine learning training samples and prediction samples.

[0021] S101, when generating the operational data set by timestamp, data rows are established using battery status data and vehicle operation data within the same acquisition period. Before establishing a data row, it is checked whether the vehicle identifier, battery object identifier, and timestamp fields correspond to the same operational object. Records that correspond to the same operational object enter the alignment process; records that do not correspond to the same operational object do not enter the current prediction processing chain. The alignment process is based on the timestamp, reading the acquisition time difference of adjacent records within the same acquisition period and comparing the acquisition time difference with a preset synchronization interval. When the preset synchronization interval is met, the corresponding data row is marked as a reserved data row; when the preset synchronization interval is not met, the corresponding data row is discarded or marked as a data row to be verified. The reserved data row serves as the main input for subsequent mining transportation operation segment division, while the data row to be verified only participates in subsequent processes after anomaly identification or replacement input processing is completed. This synchronization process provides a time consistency basis for machine learning input samples, avoiding data rows with time mismatches from directly affecting model training or inference input.

[0022] S102, perform field integrity and field continuity checks on the retained data rows and the data rows to be verified. Field integrity checks are performed item by item for voltage, current, temperature, state of charge, number of charge / discharge cycles, mileage, load, operating time, and vehicle driving status. If any necessary field is missing, the corresponding data row to be verified is marked as missing field. Field continuity checks are performed on the timestamp, mileage, state of charge, and vehicle driving status within the same continuous data collection period for the same vehicle. If there is a jump between consecutive fields that cannot be directly connected by adjacent data collection periods, the corresponding data row to be verified is marked as a field jump. The output of the field integrity and field continuity checks is a data row status label. The data row status label is passed to the abnormal data rollback processing and sample confidence weight generation stages. The data row status label is also used in the machine learning input stage to distinguish between original valid samples, samples to be verified, and substitute input samples.

[0023] S103, when a data row to be verified has missing or abruptly changed fields, the most recent data row from the same mining dump truck that meets the preset synchronization interval is read as the first alternative data source. When reading the first alternative data source, the processing object is limited to data rows that are preferentially matched from the same vehicle, the same battery object, and the same fragment type. If there is a data row in the first alternative data source with the same fragment type as the data row to be verified, the corresponding field that can fill in the currently missing field or replace the abruptly changed field is extracted from the first alternative data source, and the extracted field is written into the alternative input buffer, while the alternative source mark is retained. The alternative source mark includes the source vehicle, source collection period, source fragment type, and alternative field name, which is used to trace the alternative input source later. In the process of constructing machine learning samples, the alternative source mark and the data row status label together serve as the basis for determining the input credibility.

[0024] S104, when the first alternative data source does not contain a data row of the same fragment type as the data row to be verified, a data row of the same fragment type from historical operation fragments of the same type of mining dump truck is read as the second alternative data source. The determination of the same type of mining dump truck is based on vehicle type and mining transportation operation attributes, without introducing additional vehicle categories. After reading the second alternative data source, data rows of the same fragment type are used as candidate sources, and data rows with complete fields, continuous timestamps, and corresponding working conditions that match the data row to be verified are selected first. If there are multiple data rows that meet the conditions among the candidate sources, the selection order is fragment type consistency, field completeness, and time continuity, and the second alternative data source is formed after selection. This historical operation fragment reading process provides a traceable alternative input source for the machine learning model under the condition of incomplete data.

[0025] S105: Generate alternative inputs based on either the first or second alternative data source, and set the sample confidence weight of the alternative inputs to be lower than the sample confidence weight of the data rows that meet the preset synchronization interval. When generating alternative inputs, the processor merges the alternative fields with the non-abnormal fields in the data rows to be verified into alternative data rows. The alternative data rows retain the alternative source marker and data row status label. If neither the first nor the second alternative data source can provide fields corresponding to the current segment type, the corresponding data row to be verified will not participate in the generation of periodic aging stress features. The output of S1 includes retained data rows, data rows to be verified, alternative data rows, and a running data set. Subsequently, S2 uses the filtered data rows as the segment division object, and S4 uses the data row status label and the alternative source marker as the confidence judgment basis. During machine learning input processing, alternative data rows are not processed the same as retained data rows, but instead enter the feature sequence with a lower sample confidence weight.

[0026] S2 divides the operational data set into multiple mining transportation operation segments based on load changes, current direction, state of charge (SOC) changes, and vehicle driving status. During S2 execution, retained data rows and data rows after substitution input processing are read sequentially. The load changes, current direction, SOC trends, and vehicle driving status between adjacent data rows are used as the criteria for segment boundary determination. Load changes are used to identify loading or unloading status changes; current direction and SOC trends are used to identify charging or discharging directions; and vehicle driving status is used to identify driving or parking status. The segment division results are represented by segment type, start time, end time, and associated data row sequence, and written to the segment buffer. The segment type, start time, and end time in the segment buffer are used to form the time-series chemical condition input that the machine learning model can accept.

[0027] S201, when the load change between adjacent data rows satisfies the loading or unloading state change condition, the segment boundary is determined. When the load increases from low to high and corresponds to the waiting or low-speed state in the vehicle's driving state, the processor records the time position between adjacent data rows as the loading-related segment boundary; when the load decreases from high to low and corresponds to the waiting or low-speed state in the vehicle's driving state, the processor records the time position between adjacent data rows as the unloading-related segment boundary. After the segment boundary is formed, the data rows before the boundary are assigned to the previous segment, and the data rows after the boundary are assigned to the next segment. The segment start and end points are written to the segment buffer along with timestamps. The segment boundary makes the machine learning input no longer just a continuous numerical sequence, but a segment sequence with the semantics of mining transportation operations.

[0028] S202, when the current direction and state of charge (SOC) change trend between adjacent data rows meet the condition of being consistent with the charging or discharging direction, a charging segment or operating segment is determined. When the current direction corresponds to the charging direction and the SOC shows an upward trend, the adjacent data row is assigned to a charging segment; when the current direction corresponds to the discharging direction and the SOC shows a downward trend, and the vehicle is in a driving state, the adjacent data row is assigned to an operating segment. When the current direction and SOC change trend are inconsistent, the corresponding data row is not directly used for segment type confirmation, but is passed to the direction consistency judgment in S4. Through this process, the formation basis of charging and operating segments can be traced back to the current direction, SOC change, and vehicle driving state. This traceable segment source is used to support the machine learning model in distinguishing the input contribution of battery aging stress under different operating conditions.

[0029] S203: When the vehicle's driving state meets the parking conditions and the current is within the resting range, a parking segment is determined. The start time of the parking segment is taken from the timestamp of the first data line when the vehicle enters the parking state and the current is within the resting range, and the end time of the parking segment is taken from the timestamp of the adjacent data line when the vehicle leaves the parking state or the current leaves the resting range. If temperature, voltage, or state of charge records exist during the parking state, the corresponding records are retained in the parking segment for subsequent static voltage recovery feature extraction. The output of S2 is multiple mining transportation operation segments with segment type, start time, end time, and associated data line sequence. After the parking segments are formed, the machine learning model can retain the source of battery recoverability features in the non-driving state at the input level.

[0030] S3 extracts current continuity features, temperature range features, state of charge (SOC) swing features, load continuity features, energy consumption per unit mileage features, depth of charge features, and resting voltage recovery features from mining transportation operation segments to generate periodic aging stress features. During S3 execution, the processing program reads data rows within each mining transportation operation segment by segment type, statistically analyzes and organizes relevant fields such as current, temperature, SOC, load, mileage, and SOC to form segment-level aging stress features. These segment-level aging stress features are then combined according to segment type and time sequence within the same operation cycle to output periodic aging stress features. Periodic aging stress features are the main feature input of the prediction model containing LSTM hidden layers, used to express the comprehensive impact of different segments on battery life changes within the operation cycle.

[0031] S301, according to the segment type of mining transportation operation segments, statistical analysis is performed on the current duration interval, temperature duration interval, state of charge variation range, load duration, mileage variation, and depth of charge. The current duration interval is formed by the current direction and continuous current recording within the segment; the temperature duration interval is formed by the continuous state of temperature recording within the segment; the state of charge variation range is formed by the state of charge variation in the segment's start and end data rows; the load duration is formed by the continuous recording of the load field within the segment; the mileage variation is formed by the mileage difference in the segment's start and end data rows; and the depth of charge is formed by the state of charge variation range within the charging segment. Each statistical result is written into the segment-level aging stress characteristics, along with segment type and time sequence labels. The segment type and time sequence labels enable machine learning training samples to retain the feature source location.

[0032] S302, when a mining transportation operation segment is a running segment, the change in energy consumption per unit mileage and the duration of the load are written into the segment-level aging stress features; when a mining transportation operation segment is a charging segment, the charging depth and temperature duration range are written into the segment-level aging stress features; when a mining transportation operation segment is a parking segment, the static voltage recovery feature is written into the segment-level aging stress features. The change in energy consumption per unit mileage in the running segment is formed by the correspondence between the changes in current, state of charge, and mileage; the charging depth and temperature duration range in the charging segment represent the operating status during charging; the static voltage recovery feature in the parking segment is formed by the voltage record changes during the start and end of the parking segment. Different segment types correspond to different field combinations, avoiding the use of fields from charging, running, and parking processes in the same way for subsequent predictions. This segment-based feature organization method provides hierarchical input for machine learning models, reducing the mixing of features from different operating states in a unified input.

[0033] S303 generates periodic aging stress features based on the segment type and time sequence within the same work cycle according to the segment-level aging stress features. Within the same work cycle, the processing program arranges the corresponding segment-level aging stress features into periodic feature records according to the segment type of the work process, such as loading waiting, heavy-load driving, unloading waiting, empty return, charging, and parking. The periodic feature record retains the segment source, field source, segment start time, segment end time, and data row status label. The output of S3 is the periodic aging stress feature that can be called by S5, and the segment source relationship is passed to the subsequent feedback correction process. When the machine learning model calls the periodic feature record, it can read the segment source and field source, thus maintaining the correspondence between the predicted output and the working condition segment.

[0034] S4 generates sample confidence weights based on the temporal, directional, and state consistency between battery state data and vehicle operation data. During S4 execution, the program determines the continuity of timestamps, the correspondence of charging / discharging directions, and the matching of operating conditions within the same data row or segment. Based on the determination results, it determines the weight state for data row or segment-level features when participating in training or inference. This weight state is used to control sample participation in machine learning training and inference. The sample confidence weights do not change the original field values ​​but represent the degree of participation of data row, segment-level features, or periodic aging stress features in subsequent model inputs.

[0035] In step S401, when the timestamps of battery status data and vehicle operation data are continuous and the acquisition time difference meets the preset synchronization interval, they are determined to be time-consistent. The input for time consistency determination is the timestamp within the same acquisition cycle, the acquisition time difference, and the data row status label; the output is a time-consistent state or a time-inconsistent state. The corresponding data row of the time-inconsistent state can continue to be retained in the running data set, but it will not be used as a high-confidence input when generating sample confidence weights later. The time-consistent state is passed to step S404, where it, together with the direction-consistent state and the state-consistent state, determines the weight type. Before receiving input, the machine learning model filters or reduces the weight of time-mismatched samples through the time-consistent state.

[0036] S402, when the direction of the change in state of charge matches the charging / discharging direction corresponding to the current direction, it is determined to be consistent in direction. The consistency determination reads the trend of the change in state of charge and the current direction field within the same segment. When the state of charge increases and the current direction corresponds to the charging direction, a consistent direction is output; when the state of charge decreases and the current direction corresponds to the discharging direction, a consistent direction is output; when the direction of the change in state of charge and the current direction do not correspond, an inconsistent direction is output. The inconsistent direction state serves as a trigger condition for reducing the sample's credibility weight and is retained in the corresponding mine transportation operation segment. The consistent direction state provides a basis for verifying the charging / discharging semantics of the machine learning model input.

[0037] S403: When the load change, vehicle driving status, and current duration interval meet the preset working condition matching relationship, the state is determined to be consistent. State consistency determination is based on segment type, reading the load field, vehicle driving status field, and current duration interval. In a running segment, if the load duration record, vehicle driving status, and current duration interval correspond to the same transportation working condition, a consistent state is output; in a charging segment, if the vehicle driving status, state of charge change, and current direction correspond to the charging process, a consistent state is output; in a parking segment, if the vehicle driving status and current resting interval correspond to the parking process, a consistent state is output. If no correspondence can be formed, an inconsistent state is output. State consistency determination allows machine learning samples to not only rely on a single battery field but also combine vehicle operating status to confirm the source of the working condition.

[0038] In S404, a first sample confidence weight is generated when time, direction, and state are all consistent. If any one of these conditions is not met, a second sample confidence weight lower than the first sample confidence weight is generated and written into the corresponding mine transportation operation segment. If the input source is a substitute input, even if time, direction, and state are all consistent, the sample confidence weight corresponding to the substitute input will still be lower than the original retained data row that meets the preset synchronization interval. The output of S4 is the sample confidence weight corresponding to the data row, segment-level feature, and periodic-level aging stress feature. Subsequently, S5 sets the feature participation level according to this weight. During machine learning training or inference, the first sample confidence weight and the second sample confidence weight correspond to high-confidence sample input and reduced-weight sample input, respectively.

[0039] S5 inputs the periodic aging stress features and sample confidence weights into the prediction model containing an LSTM hidden layer, and outputs the battery life state or remaining usage time prediction result. When executing S5, the processing program uses the same operating cycle as the input boundary, arranges the periodic aging stress features into a feature sequence according to segment type, start time, and end time, and binds the sample confidence weights corresponding to the feature sequence to each segment-level feature. The prediction model containing the LSTM hidden layer forms the prediction input through the segment type, aging stress features, and sample confidence weights in the feature sequence. After receiving the feature sequence with the weights set, the prediction model containing the LSTM hidden layer performs inference and outputs the battery life state, remaining usage time, or at least one of the two as the prediction result.

[0040] S501, the periodic aging stress features within the same work cycle are formed into a feature sequence according to segment type, start time, and end time. The feature sequence retains the temporal position of each segment-level aging stress feature within the work cycle. The segment type field is used to identify other transportation operation segments supported by the running segment, charging segment, and parking segment. The start time and end time are used to determine the segment boundaries. If there are segment-level features corresponding to the data row to be verified within the same work cycle, the feature sequence synchronously retains the substitute source marker or data row status label for S502 to set the participation weights. This feature sequence serves as the basis for the textual implementation of the machine learning model input vector or equivalent input record.

[0041] S502, assign participation weights to fragment-level features in the feature sequence according to the sample confidence weights corresponding to the feature sequence. When a fragment-level feature originates from a reserved data row and is consistent in time, direction, and state, the corresponding fragment-level feature adopts the first sample confidence weight. When a fragment-level feature originates from inconsistent time, direction, or state, or is a substitute input, the corresponding fragment-level feature adopts the second sample confidence weight or a weight lower than the weight corresponding to the original reserved data row. After setting the participation weights, a weighted input record is formed and passed to S503. The weighted input record is used to distinguish the participation degree of different fragment-level features during machine learning inference.

[0042] S503, the input to the prediction model inference process, which includes an LSTM hidden layer, is a weighted input record. The output is the prediction result corresponding to the same vehicle, the same battery object, and the same work cycle. After the output is completed, the prediction result is written to the prediction result record area. The prediction result record area retains at least the prediction output time, work cycle, work segment type, source of the segment involved in the prediction, and corresponding sample confidence weight, so that S6 can establish the correspondence between the previous output and subsequent records. The machine learning inference output and the segment source are stored together, so that subsequent residual correction can trace back to the working condition input on which the model prediction is based.

[0043] S6 generates prediction residuals based on the differences between subsequent operational data or battery detection records and the prediction results, and updates the correction coefficients or input weights in subsequent predictions according to the job segment type corresponding to the prediction residuals. When S6 is executed, the processor reads the prediction output time, job cycle, and job segment type from the prediction result recording area, and then receives subsequent operational data or battery detection records generated by maintenance and inspection, determining whether the recording time falls within the subsequent correction cycle corresponding to the previous prediction result. If it falls within the subsequent correction cycle, a correspondence is established between the previous prediction result and the subsequent record; if it does not fall within the subsequent correction cycle, the subsequent record does not participate in this residual generation. S6 enables the subsequent inputs of the prediction model containing the LSTM hidden layer to be subject to operational feedback constraints.

[0044] S601, when generating prediction residuals, a correspondence is established between the prediction output time, operation cycle, and operation segment type corresponding to the prediction result and subsequent operating data or battery detection records. This correspondence is defined by the same vehicle, the same battery object, and subsequent correction cycles. After the processing program confirms that the timestamp, vehicle identifier, and battery object identifier of subsequent records match the previous prediction result record area, it reads the record fields that represent the lifespan status from the subsequent records. These record fields serve as input for the subsequent lifespan status and participate in the difference formation process. This correspondence is used in the machine learning feedback correction stage to confirm the source of the residual.

[0045] S602: When the recording time of subsequent operating data or battery detection records falls within the subsequent correction cycle corresponding to the prediction result, calculate the difference between the prediction result and the lifespan state in the subsequent operating data or battery detection records. The difference represents the offset direction and offset state between the previous output and the subsequent record. If the subsequent record is missing, the timestamp is discontinuous, or the vehicle identifier or battery object identifier cannot be matched, no difference is generated, and the record is returned as not participating in the correction. When a difference can be generated, the difference is bound to the corresponding job segment type and output to the residual record area. The difference serves as the correction basis for the machine learning feedback samples.

[0046] S603, update the correction coefficients or input weights in subsequent predictions according to the job segment type corresponding to the difference. The program reads the residual records of historical job segments of the same type and compares the offset direction of the residual records of historical job segments of the same type with that of the current prediction residual. When the offset direction of the residual records of historical job segments of the same type is consistent with that of the prediction residual, increase the correction coefficient of the corresponding job segment type in subsequent predictions; when the offset direction of the residual records of historical job segments of the same type is inconsistent with that of the prediction residual, decrease the input weight of the corresponding job segment type in subsequent predictions or keep the input weight unchanged. The updated correction coefficients or input weights are written to the subsequent prediction configuration area. The updated correction coefficients or input weights are used as input adjustment conditions for the machine learning model in subsequent predictions.

[0047] In S604, upon initiating subsequent predictions, the updated correction coefficients or input weights are read from the subsequent prediction configuration area and, together with the periodic aging stress characteristics and sample confidence weights formed in the new work cycle, are fed into S5. This closed loop establishes a traceable feedback relationship between the previous prediction results, subsequent operational data, battery detection records, and work segment types. During verification, the prediction output time, work cycle, work segment type, difference, correction coefficients, and input weights in the prediction result record area, residual record area, and subsequent prediction configuration area can be reviewed to confirm that each correction has a source record and segment attribution. Thus, the machine learning prediction process can receive feedback correction inputs with clear sources in subsequent cycles.

[0048] Through the workflow of Example 1, continuously collected battery and vehicle operation records are converted into fragmented inputs corresponding to the mining transportation operation process. Fragment-level aging stress characteristics are transformed into periodic-level aging stress characteristics according to the operation cycle, and then the data source status is distinguished by sample confidence weights. After subsequent records form prediction residuals with previous outputs, they can be written back to the correction coefficients or input weights of the corresponding operation fragment type. Throughout the entire process, the input source, processing object, intermediate state, output result, and feedback entry can all be traced through data rows, fragment buffer areas, prediction result recording areas, and residual recording areas.

[0049] Example 2 Please see Figure 1 , Figures 4 to 6 This embodiment provides a battery life prediction system for mining dump trucks based on operational data. The difference between this embodiment and Embodiment 1 is that this embodiment uses a system unit collaborative approach to implement the method flow of Embodiment 1. It completes the processing chain from operational record access to prediction feedback correction through data transfer between the operational data access unit, the operational condition segment generation unit, the aging stress feature generation unit, the sample credibility processing unit, the life prediction processing unit, and the prediction feedback correction unit. In this embodiment, the operational data access unit receives battery status data and vehicle operation data and writes them to a unified storage area according to timestamps; the operational condition segment generation unit reads the operational data set from the unified storage area and generates mining transportation operation segments; the aging stress feature generation unit extracts segment features from the mining transportation operation segments and forms periodic aging stress features; the sample credibility processing unit generates sample credibility weights based on the time consistency, direction consistency, and state consistency between battery status data and vehicle operation data; and the prediction feedback correction unit updates the correction coefficients or input weights based on the differences between subsequent operational data or battery detection records and the prediction results. The machine learning processing in this embodiment is called by the life prediction processing unit and is subject to the input constraints of the aging stress feature generation unit, the sample credibility processing unit, and the prediction feedback correction unit.

[0050] The input terminal of the operational data access unit receives voltage, current, temperature, state of charge, number of charge / discharge cycles, mileage, load, operation duration, and vehicle driving status. The output terminal writes data rows with timestamp alignment completed within the same acquisition cycle into a unified storage area. If the acquisition time difference meets the preset synchronization interval, the operational data access unit marks the data row as a reserved data row; if the acquisition time difference does not meet the preset synchronization interval, the operational data access unit marks the data row as a data row to be verified or does not write it into the subsequent processing queue. The operational condition segment generation unit reads the reserved data rows and the data rows processed by substitution input from the unified storage area. Based on load changes, current direction, state of charge changes, and vehicle driving status, it generates segment type, start time, end time, and associated data row sequence, and outputs the formed mine transportation operation segment to the aging stress feature generation unit and the sample credibility processing unit. The data row status labels in the unified storage area provide a source for data quality differentiation before machine learning model input.

[0051] After receiving mining transportation operation segments, the aging stress feature generation unit statistically analyzes the current duration interval, temperature duration interval, state of charge change range, load duration, mileage change, and depth of charge according to segment type. The unit mileage energy consumption change and load duration corresponding to the operating segment, the depth of charge and temperature duration interval corresponding to the charging segment, and the static voltage recovery characteristics corresponding to the parked segment are written into segment-level aging stress features. These segment-level aging stress features are combined according to segment type and time sequence within the same operation cycle to form period-level aging stress features, which are then transmitted to the lifetime prediction processing unit. The sample credibility processing unit synchronously reads the timestamp, acquisition time difference, state of charge change direction, current direction, load change, vehicle driving status, and current duration interval between battery status data and vehicle operation data, forming time-consistent, direction-consistent, and state-consistent states. The first sample credibility weight or the second sample credibility weight is generated from these three consistent states. Therefore, the machine learning input received by the lifetime prediction processing unit simultaneously includes feature values, segment source, and credibility weight.

[0052] After receiving the periodic aging stress features and sample confidence weights, the lifetime prediction processing unit forms a feature sequence from the periodic aging stress features within the same work cycle according to segment type, start time, and end time. Then, it sets the participation weights of the segment-level features according to the corresponding sample confidence weights. The feature sequence with the set participation weights is input into the prediction model containing the LSTM hidden layer. The model outputs the battery life status or remaining usage time prediction result and writes the prediction result record area. The prediction feedback correction unit reads the prediction output time, work cycle, and work segment type from the prediction result record area and receives subsequent running data or battery detection records. When a subsequent record falls into the corresponding subsequent correction cycle, the prediction feedback correction unit generates a difference and prediction residual. Then, according to the work segment type corresponding to the prediction residual, it reads the residual records of historical work segments of the same type and updates the correction coefficients or input weights. The updated result is returned to the lifetime prediction processing unit as the input configuration for the next round of prediction. This return path allows the prediction model containing the LSTM hidden layer to receive the feedback-corrected input configuration in the next round of inference.

[0053] In the anomaly handling collaboration, the data access unit sends field missing status or field transition status to the sample credibility processing unit and the lifetime prediction processing unit. The sample credibility processing unit first reads the data row from the most recent operation of the same mining dump truck that met the preset synchronization interval as the first alternative data source. If the first alternative data source does not contain a data row of the same fragment type as the data row to be verified, it reads the data row of the same fragment type from the historical operation fragments of the same type of mining dump truck as the second alternative data source. After forming the alternative input from the first or second alternative data source, the system sets the sample credibility weight of the alternative input to be lower than the sample credibility weight corresponding to the data row that met the preset synchronization interval. The anomaly handling output includes an alternative source marker, an alternative field name, and a corresponding weight status, which subsequent modules can use to distinguish between the original retained data and the alternative input data. After reading this weight status during the machine learning processing stage, the alternative input can be used as a low-credibility sample in the prediction input.

[0054] System verification in Example 2 can be performed through unit input / output states. The running data access unit outputs retained data rows, data rows to be verified, and a set of running data; the working condition segment generation unit outputs mining transportation operation segments with segment type, start time, and end time; the aging stress feature generation unit outputs periodic aging stress features; the sample credibility processing unit outputs sample credibility weights; the lifetime prediction processing unit outputs prediction results; and the prediction feedback correction unit outputs prediction residuals and updated correction coefficients or input weights. Each output has a clearly defined receiving unit and subsequent call location, forming a data flow closed loop corresponding to Example 1. The machine learning inference performed by the lifetime prediction processing unit can be traced and verified through the input feature sequence, sample credibility weights, and prediction result recording area.

[0055] The foregoing embodiments are used to illustrate the implementable processes, data objects, judgment conditions, anomaly handling, unit collaboration, and feedback correction relationships of the present invention. The data fields, processing order, fragment types, weight states, model inputs, output records, and reference relationships in the accompanying drawings listed in the embodiments are used to explain the technical solutions described in the application documents and should not limit the present invention to the specific embodiments listed. The machine learning-related content is only used to illustrate the relationship between the training sample construction, feature input, inference output, and feedback correction of the lifetime prediction model in the technical chain of the present invention. Without departing from the scope of the technical solutions described in the present invention, any equivalent substitutions, equivalent modifications, equivalent combinations, order adjustments, module correspondence replacements, equivalent transformations of field names, equivalent inheritance of execution subjects, or equivalent changes in carrier forms that can be conceived by those skilled in the art should fall within the scope of protection of this patent; however, it should not be extended to unclaimed topics, nor should the substantive correspondence between technical objects be changed through name changes.

Claims

1. A method for predicting the battery life of mining dump trucks based on operational data, characterized in that, Acquire battery status data and vehicle operation data of mining dump trucks, and generate an operation data set according to timestamps; Based on changes in load, current direction, state of charge, and vehicle driving status, the operational data set is divided into multiple mining transportation operation segments. Extract current continuity characteristics, temperature range characteristics, state of charge swing characteristics, load continuity characteristics, energy consumption per unit mileage characteristics, charging depth characteristics, and static voltage recovery characteristics from the mining transportation operation segment to generate periodic aging stress characteristics. When the battery status data and the vehicle operation data are consistent in time, direction and status, the sample confidence weight is optimized; when any consistency condition is not met, the sample confidence weight is reduced. The cycle-level aging stress characteristics and the sample confidence weights are input into a prediction model containing an LSTM hidden layer, and the battery life status or remaining usage time prediction results are output. Based on the difference between subsequent operational data or battery detection records and the prediction results, a prediction residual is generated, and the correction coefficients or input weights in subsequent predictions are updated according to the job segment type corresponding to the prediction residual.

2. The method for predicting the battery life of mining dump trucks based on operational data according to claim 1, characterized in that, The battery status data includes voltage, current, temperature, state of charge, and number of charge / discharge cycles; the vehicle operation data includes mileage, load, operation time, and vehicle driving status. When generating the running data set according to the timestamp, data rows are established using the battery status data and the vehicle operation data within the same collection period. When the collection time difference between the battery status data and the vehicle operation data meets the preset synchronization interval, the data row is retained. When the collection time difference does not meet the preset synchronization interval, the corresponding data row is discarded or marked as a data row to be verified.

3. The method for predicting the battery life of mining dump trucks based on operational data according to claim 2, characterized in that, When dividing the operational data set into multiple mining transportation operation segments, the retained data rows and the data rows to be verified are read in chronological order. When the load change between adjacent data rows meets the loading or unloading state change conditions, the segment boundary is determined. When the current direction between adjacent data rows and the state of charge change trend meet the charging or discharging direction conditions, the charging segment or running segment is determined. When the vehicle driving state meets the parking conditions and the current is in the static range, the parking segment is determined. The segment type, start time, and end time of the mining transportation operation segment are generated based on the segment boundary, the charging segment, the running segment, and the parking segment.

4. The method for predicting the battery life of mining dump trucks based on operational data according to claim 1, characterized in that, When generating the periodic aging stress characteristics, the current duration interval, temperature duration interval, state of charge variation range, load duration, mileage variation and depth of charge are statistically analyzed according to the segment type of the mining transportation operation segment. When the mining transportation operation segment is a running segment, the change in energy consumption per unit mileage and the duration of load are written into the segment-level aging stress characteristics. When the mining transportation operation segment is a charging segment, the charging depth and temperature duration range are written into the segment-level aging stress characteristics. When the mining transportation operation segment is a parking segment, the static voltage recovery characteristics are written into the segment-level aging stress characteristics. The cycle-level aging stress characteristics are generated based on the segment-level aging stress characteristics within the same working cycle, according to the segment type and time sequence.

5. The method for predicting the battery life of mining dump trucks based on operational data according to claim 1, characterized in that, When generating the sample confidence weight, the battery status data and the vehicle operation data are considered to be in time consistency when their timestamps are continuous and the time difference between their collection meets the preset synchronization interval. When the direction of the change in the state of charge matches the charging and discharging direction corresponding to the current direction, it is determined that the directions are consistent. When the load change, vehicle driving status and current duration range meet the preset working condition matching relationship, it is determined that the state is consistent. When the time, direction, and state are all consistent, a first sample confidence weight is generated. When any one of the determinations is not true, a second sample confidence weight lower than the first sample confidence weight is generated, and the second sample confidence weight is written into the corresponding mine transportation operation segment.

6. The method for predicting the battery life of mining dump trucks based on operational data according to claim 1, characterized in that, When the periodic aging stress features and the sample confidence weights are input into a prediction model containing an LSTM hidden layer, the periodic aging stress features within the same work cycle are formed into a feature sequence according to the segment type, start time, and end time. The fragment-level features in the feature sequence are assigned participation weights according to the sample confidence weights corresponding to the feature sequence. The feature sequence with weighted participation is input into the prediction model containing the LSTM hidden layer to obtain at least one of the battery life state and remaining usage time as the prediction result.

7. The method for predicting the battery life of a mining dump truck based on operational data according to claim 6, characterized in that, When generating the prediction residual, a correspondence is established between the prediction output time, operation cycle and operation segment type corresponding to the prediction result and subsequent operation data or battery detection records; When the recording time of the subsequent operating data or the battery detection record falls within the subsequent correction period corresponding to the prediction result, the difference between the prediction result and the life state in the subsequent operating data or the battery detection record is calculated. The correction coefficients or input weights in subsequent predictions are updated according to the job segment type corresponding to the difference.

8. The method for predicting the battery life of a mining dump truck based on operational data according to claim 1, characterized in that, When updating the correction coefficient or the input weight, the residual records of historical job segments of the same type are read according to the job segment type corresponding to the predicted residual; When the residual record of the same type of historical job segment is consistent with the offset direction of the predicted residual, the correction coefficient of the corresponding job segment type in the subsequent prediction is optimized; When the offset direction of the residual record of the same type of historical job segment is inconsistent with that of the predicted residual, the input weight of the corresponding job segment type in the subsequent prediction is reduced or the input weight is kept unchanged.

9. The method for predicting the battery life of a mining dump truck based on operational data according to claim 2, characterized in that, When the data row to be verified has missing fields or fields that change, the data row of the same mining dump truck that most recently met the preset synchronization interval is read as the first alternative data source. When the first alternative data source does not have a data row of the same fragment type as the data row to be verified, the data row of the same fragment type in the historical operation fragment of the same type of mining dump truck is read as the second alternative data source. An alternative input is generated based on the first alternative data source or the second alternative data source, and the sample confidence weight of the alternative input is set to be lower than the sample confidence weight corresponding to the data line that meets the preset synchronization interval.

10. A battery life prediction system for mining dump trucks based on operational data, used to support the battery life prediction method for mining dump trucks as defined in any one of claims 1 to 9, characterized in that, It includes an operational data access unit, a working condition segment generation unit, an aging stress characteristic generation unit, a sample credibility processing unit, a life prediction processing unit, and a prediction feedback correction unit. The operation data access unit receives battery status data and vehicle operation data, and writes them to a unified storage area according to timestamps; The working condition segment generation unit generates mining transportation operation segments based on load changes, current direction, state of charge changes, and vehicle driving status. The aging stress feature generation unit extracts segment features from the mining transportation operation segment and forms periodic aging stress features; The sample credibility processing unit generates sample credibility weights based on the time consistency, direction consistency, and state consistency between the battery status data and the vehicle operation data. The life prediction processing unit inputs the cycle-level aging stress characteristics and the sample confidence weights into the prediction model containing the LSTM hidden layer, and outputs the battery life status or remaining service time prediction results. The prediction feedback correction unit updates the correction coefficients or input weights based on the differences between subsequent operating data or battery detection records and the prediction results.

Images