A method for predicting the completion rate of a sanitation vehicle operation based on machine learning

Abstract

The application discloses a kind of based on machine learning's environmental sanitation vehicle operation completion rate prediction method, it is related to machine learning technical field, including, to environmental sanitation vehicle current operation state is graded mapping, obtains operation state token, operation state token is spliced, and additional time step coding forms operation state token sequence;Operation state token sequence is input into conditional sequence machine learning model, generates the operation completion rate prediction value of current operation stage;By binding evidence abstract and operation completion rate prediction value, obtain completion rate prediction result package;According to completion rate prediction result package, the reasoning process of conditional sequence machine learning model is consistent correction, and the corrected operation completion rate prediction value is output.The application realizes the stable prediction and explainable reasoning of environmental sanitation vehicle operation completion rate by constructing layered semantic clear operation state token system and introducing conditional sequence machine learning model with causal constraint.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a machine learning-based method for predicting the completion rate of sanitation vehicle operations. Background Technology

[0002] In the context of smart sanitation and refined urban management, the dynamic assessment and prediction of sanitation vehicle operation completion rates has gradually become a key technical issue in intelligent scheduling and operation supervision. Existing research typically uses rule-driven or statistical learning methods to model the operational status of sanitation vehicles. For example, it utilizes indicators such as the percentage of work completed, vehicle load margin, and distance to work nodes to construct empirical formulas or regression models to estimate work progress or completion status. In recent years, with the development of machine learning technology, some solutions have begun to introduce time series models or deep learning methods to model historical operation data to characterize the patterns of operation status changes over time, thereby improving the accuracy and automation level of completion rate prediction.

[0003] However, existing models often simply concatenate or equally weight the characteristics of the work status, lacking an effective characterization of the causal hierarchy among various factors such as work target constraints, load matching relationships, and spatial accessibility. This results in insufficient stability of the prediction results under complex working conditions. On the other hand, most prediction models only output a single numerical result, making it difficult to explain the basis for the completion rate prediction and unable to perform consistency correction for abnormal fluctuations in the prediction process. This limits the credibility and engineering applicability of the prediction results in scheduling decision-making and supervision scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a machine learning-based method for predicting the completion rate of sanitation vehicle operations to address the problem of insufficient stability of prediction results under complex working conditions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a machine learning-based method for predicting the completion rate of sanitation vehicle operations, comprising, The current operating status of sanitation vehicles is hierarchically mapped to obtain operating status tokens. The operating status tokens are then concatenated and time step codes are added to form an operating status token sequence. The operation status token sequence is input into the conditional sequence machine learning model, and the completion state perception reasoning mechanism is used to reason about the evolution trend of sanitation vehicles from operation state to operation completion state, and generate the operation completion rate prediction value of the current operation stage. The aggregated conditional sequence machine learning model focuses on the changes in the intensity of attention to the job status token sequence during the inference process, forming an evidence sequence. The evidence sequence is then compressed to obtain an evidence summary. By binding the evidence summary with the job completion rate prediction value, a completion rate prediction result package is obtained. Based on the completion rate prediction results, the inference process of the conditional sequence machine learning model is corrected for consistency, and the corrected job completion rate prediction value is output.

[0007] As a preferred embodiment of the machine learning-based sanitation vehicle operation completion rate prediction method of the present invention, the operation status token includes a target completion rate instruction token, a node compatibility perception token, and a visibility token.

[0008] As a preferred embodiment of the machine learning-based sanitation vehicle operation completion rate prediction method of the present invention, the operation status token concatenation means that, according to the causal hierarchy of the predicted operation completion rate, the target completion rate instruction token is used as the global inference constraint as the priority input, and the subsequent node is compatible with perception tokens and visibility tokens.

[0009] As a preferred embodiment of the machine learning-based method for predicting the completion rate of sanitation vehicle operations according to the present invention, the specific steps for inferring the evolution trend of sanitation vehicles from the operational state to the operational completed state and generating a predicted value for the completion rate of the current operational stage are as follows. Using the job status token sequence as input, a conditional sequence machine learning model is constructed based on an autoregressive sequence modeling structure. The conditional sequence machine learning model is trained using the historical job status token sequence as training samples. Input the current job status token sequence into the trained conditional sequence machine learning model to infer the job status from the job status to the job completion status, and output the predicted job completion rate of the sanitation vehicle at the current time and job status.

[0010] As a preferred embodiment of the machine learning-based method for predicting the completion rate of sanitation vehicle operations according to the present invention, the specific steps for constructing the conditional sequence machine learning model are as follows: In the autoregressive sequence modeling structure, a causal mask is set to limit the temporal correlation range of the job status token sequence, thereby solidifying the temporal dependency constraints of the autoregressive sequence modeling structure into the structural inference rules of the conditional sequence machine learning model. The autoregressive sequence modeling structure performs time-step computation on the job state token sequence under causal mask constraints to obtain the intermediate implicit representation; The intermediate implicit representation is used as the input to the completion state perception reasoning mechanism to evaluate the evolution trend from the task state to the task completion state and generate a completion confidence representation. The conditional sequence machine learning model transforms the completion confidence representation into the current job completion rate prediction at each time step, thus completing the construction of the conditional sequence machine learning model.

[0011] As a preferred embodiment of the machine learning-based method for predicting the completion rate of sanitation vehicle operations described in this invention, the evaluation of the evolution trend from the operation state to the operation completion state refers to performing linear transformation and nonlinear mapping on the intermediate implicit representation to output a completion confidence representation.

[0012] As a preferred embodiment of the machine learning-based sanitation vehicle operation completion rate prediction method of the present invention, the completion confidence representation refers to a quantitative indicator of the degree of credibility of the current operation status continuously reaching the operation completion state.

[0013] As a preferred embodiment of the machine learning-based method for predicting the completion rate of sanitation vehicle operations according to the present invention, the specific steps for binding the evidence summary with the predicted operation completion rate are as follows: Time-align the evidence summary with the corresponding predicted task completion rate; The predicted task completion rate is used as the main prediction result identifier, and the evidence summary is used as the explanatory descriptive information on which the predicted task completion rate is formed, thus establishing a correspondence between the predicted task completion rate and the evidence summary. For each job completion rate prediction, the evidence digests of each job status token are associated and bound according to the correspondence between the job completion rate prediction and the evidence digest, to obtain the completion rate prediction result package.

[0014] As a preferred embodiment of the machine learning-based method for predicting the completion rate of sanitation vehicle operations according to the present invention, the consistency correction step is as follows: Determine whether the predicted task completion rate deviates from its value across consecutive time steps; When deviations exist, the intermediate implicit representations formed during the inference process of the conditional sequence machine learning model are reweighted based on the relative contribution ratio of the main influencing factor categories in the completion rate prediction result package to the predicted value of the task completion rate. Conversely, if there is no deviation, there is no need for consistency correction.

[0015] As a preferred embodiment of the machine learning-based method for predicting the completion rate of sanitation vehicle operations described in this invention, the reweighting process refers to scaling the internal attention weights of the corresponding major influencing factor categories component by component using the normalized value of the relative contribution ratio.

[0016] The beneficial effects of this invention are as follows: By orderly concatenating the target completion rate instruction token, node compatibility perception token, and visibility token according to the causal logic of task completion, semantic confusion of multi-source state information in the model input stage is effectively avoided, and the model's ability to understand the evolution logic of task completion is improved. With the help of autoregressive modeling structure and causal mask constraints, the model reasoning process strictly follows the temporal causal relationship, enhancing the consistency and robustness of completion rate prediction in the continuous task stage. By introducing a completion state perception reasoning mechanism and reasoning attention information output, an explicit correlation between the prediction result and changes in load affordability and spatial accessibility is realized. On this basis, an evidence summary and consistency correction mechanism are constructed, thereby adaptively correcting the prediction deviation without changing the actual task behavior. Attached Figure Description

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

[0018] Figure 1 This is a flowchart of a machine learning-based method for predicting the completion rate of sanitation vehicle operations.

[0019] Figure 2 This is a flowchart of the inference process of a conditional sequence machine learning model.

[0020] Figure 3 Create flowcharts for the evidence sequence and evidence summary.

[0021] Figure 4 This is a flowchart of the consistency correction process. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a machine learning-based method for predicting the completion rate of sanitation vehicle operations, comprising the following steps: S1. Map the current operating status of sanitation vehicles in a hierarchical manner to obtain operating status tokens. Concatenate the operating status tokens and add time step codes to form an operating status token sequence.

[0026] The job status tokens include target completion rate instruction tokens, node compatibility awareness tokens, and visibility tokens.

[0027] Operation status refers to the set of states that can be objectively obtained and used to describe the operating conditions of sanitation vehicles at the current operation moment. These include assessment requirements, operation type, historical records of the completion level of similar tasks, the current remaining load of sanitation vehicles, the set of candidate operation nodes, the regular workload of each candidate operation node, the workload fluctuation of each candidate operation node, the current location of sanitation vehicles, and the spatial location of each candidate operation node.

[0028] A pre-defined numerical range with business semantics is used to map continuous values ​​obtained from different operational states of sanitation vehicles into the corresponding range, thus completing the hierarchical mapping of the current operational state of sanitation vehicles. The discrete levels obtained from this hierarchical mapping are then encoded to obtain operational state tokens. The specific steps are as follows. The target completion rate instruction token is determined based on the sanitation assessment requirements and job type, and is used to represent the expected completion level of the current job process. Specifically, based on the job type and assessment requirements, the maximum value in the historical target completion level records of similar tasks for sanitation vehicles is used as the target completion rate value. Fixed segments of the target completion rate value are determined according to the quantiles of the historical completion rate distribution of similar jobs, and the target completion rate value is discretized according to the fixed segments to obtain the target level of the target completion rate value. The target level includes multiple different levels of job completion rate target requirements, and the higher the target level, the higher the job completion rate target. The higher the requirement, for example, a target completion rate value ≥ 0.95 corresponds to target level 5, 0.90 ≤ target completion rate value < 0.95 corresponds to target level 4, 0.85 ≤ target completion rate value < 0.90 corresponds to target level 3, 0.80 ≤ target completion rate value < 0.85 corresponds to target level 2, and target completion rate value < 0.80 corresponds to target level 1. The target level of the target completion rate value is encoded in the form of "TCMD: target level" to obtain the target completion rate instruction token, which is placed at the beginning of the job status token sequence input as a conditional instruction header.

[0029] The node compatibility awareness token is determined based on the matching relationship between the current operating load status of the sanitation vehicle and the amount of work involved in the current operating stage, and is used to characterize the degree to which the operating load supports the continuous completion of the operation. Specifically, a compatibility score is calculated for each candidate operating node based on the current remaining load of the sanitation vehicle, the regular operating load of each candidate operating node, and the operating fluctuation of each candidate operating node, as shown in the following formula: ; In the formula, A i Let D be the compatibility score of the i-th candidate job node, and B be the current remaining load capacity of the sanitation vehicle. i Let C be the job float of the i-th candidate job node. i Let ε be the normal workload of the i-th candidate job node, and let ε be a very small constant, such as 0.01, to avoid division by zero; Based on the matching relationship between remaining loading capacity and workload in sanitation operations, a preset compatibility scoring threshold is established. The compatibility score is then discretized into a compatibility level according to this threshold. The compatibility level represents the ability of the current remaining loading capacity to cover the workload of all subsequent candidate work nodes. A higher compatibility level indicates a lower risk of completing the task in one go. For example, A... i ≥1 corresponds to a compatibility level of 5, and 0.5≤A i <1 corresponds to a compatibility level of 4, 0≤A i A value < 0.5 corresponds to a compatibility level of 3, and -0.5 ≤ A. i <0 corresponds to a compatibility level of 2, A.i The compatibility level corresponding to <-0.5 is level 1. The compatibility level is encoded in the form of "COMP:node number:compatibility level" to obtain the node compatibility awareness token.

[0030] The visibility token is determined based on the spatial proximity between the current work location and the remaining work area, and is used to characterize the spatial difficulty of the work process. Specifically, using the quantile discretization method, the shortest route length from the current location of the sanitation vehicle to the spatial locations of each subsequent candidate work node is statistically analyzed. Based on the statistical distribution of the historical transfer distances (transfer distances from one work node to the next) of sanitation vehicles of the same work type, discretization intervals for the shortest route length are set, such as 20%, 40%, 60%, and 80%. Discretizing the shortest route length within these intervals yields the visibility level, representing the evolution of the candidate work node's contribution to the work completion rate in terms of spatial location. Support strengths, for example, the visibility level is level 5 for historical transition distances with shortest path cost ≤ 80th percentile, level 4 for 60% ≤ shortest path cost < 80%, level 3 for 40% ≤ shortest path cost < 60%, level 2 for 20% ≤ shortest path cost < 40%, and level 1 for historical transition distances with shortest path cost < 40%. The visibility level is encoded in the form "VIS:node number:visibility level", and the visibility levels of all subsequent candidate job nodes in the candidate job node set are sorted in ascending order by node number, and a visibility token is output.

[0031] The target completion rate instruction token, node compatibility awareness token, and visibility token are concatenated and a time step code "TS:t" is added to form a job status token sequence for a single time step.

[0032] It should be noted that by sequentially concatenating the job status tokens according to the causal hierarchy of job completion rate prediction, the conditional sequence machine learning model can correctly distinguish between the prediction target, job capability constraints, and spatial reachability without introducing additional rules, thereby improving the stability and consistency of completion rate prediction. Among them, the target completion rate instruction token is given priority as the global inference constraint input, followed by the input of status information reflecting the load matching status (node ​​compatibility awareness token) and spatial reachability (visibility token). This order ensures that the conditional sequence machine learning model can infer according to the actual logic of job completion rate formation.

[0033] S2. Input the operation status token sequence into the conditional sequence machine learning model, and use the completion state perception reasoning mechanism to reason about the evolution trend of sanitation vehicles from operation state to operation completion state, and generate the operation completion rate prediction value of the current operation stage.

[0034] Taking the job status token sequence as input, a conditional sequence machine learning model is constructed based on an autoregressive sequence modeling structure. The specific steps are as follows. The conditional sequence machine learning model includes an autoregressive sequence modeling structure, a causal mask, a completion-aware inference mechanism, and a completion rate mapping component. The autoregressive sequence modeling structure performs time-step causal modeling on the embedded job status token sequence, using only historical job status information prior to the current time step for correlation calculations at each time step. A causal mask is internally set within the autoregressive sequence modeling structure to restrict access to information from future time steps, ensuring the inference process conforms to online prediction constraints. The completion-aware inference mechanism receives the intermediate latent representation output by the autoregressive sequence modeling structure and performs a completion-state reachability assessment on the intermediate latent representation sequence to generate a completion confidence representation. The completion rate mapping component receives the completion confidence representation and maps it to the predicted job completion rate value corresponding to the current time step, as well as inference attention information.

[0035] In the autoregressive sequence modeling structure, a causal mask is set to restrict the temporal correlation range of the job status token sequence. This prevents the internal computation of any time step from accessing the token information of future time steps in the job status token sequence, thereby solidifying the temporal dependency constraint of the autoregressive sequence modeling structure into the structural inference rule of the conditional sequence machine learning model. Under the constraints of causal masking, the autoregressive sequence modeling structure calculates the intermediate implicit representation of the job status token sequence step by step. Specifically, after the job status token sequence is input into the conditional sequence machine learning model in chronological order, the conditional sequence machine learning model, under the constraints of the autoregressive sequence modeling structure and causal masking, accumulates and updates the temporal dependencies between historical job status tokens layer by layer. This forms an internal expression that comprehensively reflects the evolution characteristics of job objectives (target completion rate instruction token), load matching status (node ​​compatibility awareness token), and spatial reachability (visibility token) over time. This intermediate implicit representation is used to express the trend information of job status evolving towards completion over time.

[0036] The intermediate implicit representation is used as input to the completion state perception inference mechanism to evaluate the evolution trend of the job state towards the job completion state and generate a completion confidence representation. Specifically, at each time step, the intermediate implicit representation is used as input to the completion state perception inference mechanism to analyze the trend characteristics of the current job state continuously converging towards the completion state, forming a completion confidence representation that can quantify the credibility of the current job state continuously reaching the job completion state. In the evaluation process, the completion state perception inference mechanism does not make isolated judgments on a single job state, but rather makes a comprehensive inference based on the cost change information of the node compatibility perception token and visibility token contained in the intermediate implicit representation, and the correlation change relationship in the time dimension. The completion confidence representation reflects the degree of evolution of the current job state towards the completion state in a continuously changing form.

[0037] Among them, the completion-aware reasoning mechanism is a feedforward neural network output structure embedded in the conditional sequence machine learning model. It takes the intermediate hidden representation corresponding to the current time step as input, performs linear transformation on the intermediate hidden representation, and performs nonlinear mapping using a nonlinear activation function to output the confidence level of the current task state reaching the completion state under the given task conditions, i.e., the completion confidence representation.

[0038] At each time step, the conditional sequence machine learning model converts the completion confidence representation into the predicted value of the job completion rate at the current time. Specifically, the completion confidence representation serves as a semantic expression of the job completion trend within the conditional sequence machine learning model. In the output stage, the completion confidence representation is normalized according to the parameterized output relationship formed by the conditional sequence machine learning model after training, converting the completion confidence representation into a predicted value of the job completion rate under the current job status. This allows the predicted value of the job completion rate to change continuously with the job status and maintain semantic consistency with the completion confidence representation.

[0039] Furthermore, the conditional sequence machine learning model simultaneously outputs inference attention information during the inference process of generating completion confidence representations and job completion rate predictions. Specifically, when the conditional sequence machine learning model is constructing intermediate hidden representations and inferring completion confidence representations, it assigns attention intensity (internal attention weight) to different job status tokens. In the output stage, the conditional sequence machine learning model outputs the attention intensity in an explicit form as inference attention information, so that the inference attention information can be used to explain the basis for the formation of job completion rate predictions.

[0040] It should be noted that the conditional sequence machine learning model maps the job status token sequence into an intermediate implicit representation, a completion confidence representation, and a job completion rate prediction value through a multi-level parameterized mapping process. At the same time, it outputs inference attention information that reflects the degree of influence of each job status token during the same inference process.

[0041] The historical job status token sequence, formed in chronological order, is used as training samples to train the conditional sequence machine learning model. The specific steps are as follows. The historical job status token sequences covering multiple job cycles and different job states are collected as training samples to ensure that the conditional sequence machine learning model can learn the evolution pattern from different job states to job completion states.

[0042] During training, the historical actual completion rate of the tasks corresponding to the historical task status token sequence is used as a supervision signal. The conditional sequence machine learning model follows the autoregressive sequence modeling structure and causal mask constraints. At any time step, it only uses the historical task status token sequence before the corresponding time step for inference, simulating the inference conditions in actual online prediction.

[0043] By aligning the job progress results that have occurred in the historical job status token sequence, the conditional sequence machine learning model can gradually learn the inherent laws of job status evolution to job completion state at different job stages, thereby forming a stable intermediate implicit representation and completion state reachability assessment capability within the conditional sequence machine learning model.

[0044] Under the constraints of the completion-aware reasoning mechanism, the conditional sequence machine learning model continuously adjusts its internal parameters during training to ensure that the completion confidence representation generated from the intermediate hidden representation remains consistent with the final completion state in the historical job state token sequence.

[0045] The mean squared error between the predicted completion rate and the actual completion rate of historical tasks is calculated, and the model parameters in the conditional sequence machine learning model are updated by backpropagation based on the mean squared error.

[0046] After multiple training iterations, if the mean squared error between the predicted completion rate of the conditional sequence machine learning model and the actual completion rate of historical tasks does not decrease over multiple consecutive training cycles (e.g., 10 cycles), it is determined that the completion state perception reasoning ability of the conditional sequence machine learning model has reached a stable state. At the same time, if the conditional sequence machine learning model can continuously output consistent completion confidence representations in training samples at different task stages, and the corresponding reasoning attention information does not show irregular fluctuations in the time dimension, it is confirmed that the conditional sequence machine learning model has fully learned the inherent laws of the evolution from task state to task completion state, including the correspondence between completion confidence representations and predicted completion rate values, and training ends.

[0047] Input the current job status token sequence into the trained conditional sequence machine learning model, and output the predicted job completion rate of the sanitation vehicle at the current time and job status.

[0048] It should be noted that by introducing a conditional sequence machine learning model with completion state awareness, the task completion rate can be generated based on the temporal evolution of the task state, avoiding reliance on fixed rules or simple regression calculations, and improving the pertinence and accuracy of completion rate prediction.

[0049] S3. The aggregated conditional sequence machine learning model focuses on the changes in attention intensity of the job status token sequence during the inference process to form an evidence sequence. The evidence sequence is then compressed to obtain an evidence summary. By binding the evidence summary with the job completion rate prediction value, a completion rate prediction result package is obtained.

[0050] Based on the inference attention information output by the conditional sequence machine learning model, the attention intensity (internal attention weight) of each job status token is sorted in chronological order.

[0051] The summary node compatibility perception token's attention intensity changes over time, forming an evidence sequence to characterize the trend of load affordability changes, and the summary visibility token's attention intensity changes over time, forming an evidence sequence to characterize spatial accessibility changes.

[0052] Among them, the evidence sequence is a vector formed by arranging the attention intensity of a certain type of job status token (node ​​compatibility perception token and visibility token) in the model inference process according to the time step order.

[0053] It should be noted that the evidence sequence is used to explain the impact of node-compatible perception tokens and visibility tokens on the predicted value of job completion rate under the given target completion rate instruction token constraint. The target completion rate instruction token is a "reasoning condition" rather than a "reasoning basis", and therefore is not suitable as part of the evidence sequence for the predicted value of job completion rate.

[0054] The evidence sequences of node-compatible perception tokens and visibility tokens are compressed according to the time dimension to obtain evidence summaries. Specifically, by integrating the trend of attention intensity changes in continuous time steps (such as continuous increase, continuous decrease, or relative stability), the time dimension merging process of each evidence sequence is completed to eliminate the influence of random fluctuations. The attention intensity of multiple time steps is compressed into an evidence expression result that can represent the overall effect of the operation status token, i.e., evidence summaries (such as continuous increase, continuous decrease, or relative stability).

[0055] The evidence summary is bound to the corresponding job completion rate prediction value. Specifically, using the time step as the consistency benchmark, the evidence summary and the corresponding job completion rate prediction value are time-aligned so that the evidence summary and the job completion rate prediction value correspond to the same job time or the same prediction period. The job completion rate prediction value is used as the main prediction result identifier, and the evidence summary is used as the explanatory descriptive information on which the job completion rate prediction value is formed, so that a one-to-one correspondence is established between the evidence summary and the job completion rate prediction value. For each job completion rate prediction value, a set of evidence summaries consisting of load affordability changes and spatial accessibility changes are associated. The job completion rate prediction value and the evidence summary are combined to form a structured prediction result record, resulting in a completion rate prediction result package.

[0056] It should be noted that after the completion rate prediction result package is formed, the completion rate prediction result package is output and transmitted as a whole object. When processing the completion rate prediction result, the predicted value of the job completion rate and its corresponding evidence summary are always obtained at the same time, so as to ensure that the completion rate prediction result always maintains a consistent binding state between the prediction result and the prediction basis during use, display or further processing.

[0057] It should be noted that by constructing a chain of evidence for completion rate prediction, each completion rate prediction result can be linked to specific operational influencing factors, thereby solving the problem of "difficult-to-interpret prediction results" in the application of machine learning models for completion rate prediction, and improving the interpretability and regulatory credibility of the prediction results.

[0058] S4. Based on the completion rate prediction results package, perform consistency correction on the inference process of the conditional sequence machine learning model, and output the corrected job completion rate prediction value.

[0059] Based on the completion rate prediction result package, compare whether the predicted value of the task completion rate deviates from the long-term evolution trend reflected by the evidence summary over continuous time steps. For example, if the long-term evolution trend reflected by the evidence summary is a continuous increase, while the change of the predicted value of the task completion rate over continuous time steps is a gradual decrease or relatively stable, then there is a deviation. Conversely, if the change of the predicted value of the task completion rate over continuous time steps is also a continuous increase, then there is no deviation.

[0060] If the predicted job completion rate deviates across consecutive time steps, then consistency correction needs to be performed on the inference process of the conditional sequence machine learning model. Otherwise, if there is no deviation, consistency correction is not required. The specific steps are as follows. The evidence summary in the completion rate prediction result package is analyzed to identify the main influencing factors that cause deviations in the completion rate prediction value. These include changes in load matching corresponding to node compatibility awareness tokens and changes in spatial accessibility corresponding to visibility tokens. Specifically, the changes in attention intensity of each job status token in the evidence sequence are compared category by category to determine whether, within the time interval where the job completion rate prediction value fluctuates abnormally or shows inconsistent trends, there is a situation where the change in attention intensity of a certain job status token is significantly higher than that of other job status tokens. This completes the preliminary determination (i.e., the change in attention intensity is higher than that of other job status tokens). Furthermore, a consistency analysis is performed on the change direction of each job status token in multiple consecutive time steps. For example, it is determined whether the evidence sequence corresponding to the node compatibility awareness token continues to change in a direction unfavorable to the evolution of the completion state. When a certain job status token has both preliminary determination and consistency analysis results in multiple consecutive time steps, it can be identified as the job status token corresponding to the main influencing factor category.

[0061] Keeping the operation status token sequence unchanged, and without changing the operation behavior, operation path or scheduling strategy of sanitation vehicles, the intermediate hidden representations formed in the inference process of the conditional sequence machine learning model are reweighted based on the relative contribution ratio of each influencing factor category in the completion rate prediction result package to the predicted value of operation completion rate. That is, the internal attention weight of intermediate hidden representations that are consistent with the evolution trend of completion state is increased in the inference process, while the internal attention weight of intermediate hidden representations that are inconsistent with the evolution trend of completion state is decreased.

[0062] Specifically, to adjust the weighting of internal concerns, the normalized proportion of the relative contribution of each category of influencing factors to the predicted value of the task completion rate can be used to scale the weighting of the internal concerns for the corresponding major influencing factor categories component by component.

[0063] After the conditional sequence machine learning model has undergone consistency correction, the predicted value of the job completion rate at the current moment is regenerated.

[0064] It should be noted that by performing inference consistency correction on the completion rate prediction results, the sharp fluctuations in the prediction values ​​caused by short-term state fluctuations or local anomalies are effectively suppressed, making the completion rate prediction results smoother in the time dimension and more self-consistent in the logical level. This avoids the complex method of indirectly improving the completion rate by adjusting the work behavior or scheduling strategy, and at the same time improves the stability and practicality of the prediction results in continuous operation scenarios.

[0065] This embodiment also provides a computer device applicable to the machine learning-based method for predicting the completion rate of sanitation vehicle operations, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the machine learning-based method for predicting the completion rate of sanitation vehicle operations as proposed in the above embodiment.

[0066] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0067] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the machine learning-based method for predicting the completion rate of sanitation vehicle operations as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0068] In summary, this invention effectively avoids semantic confusion of multi-source state information during the model input stage by orderly concatenating the target completion rate instruction token, node compatibility perception token, and visibility token according to the causal logic of task completion. This improves the model's ability to understand the evolutionary logic of task completion. By leveraging the autoregressive modeling structure and causal mask constraints, the model's inference process strictly follows the temporal causal relationship, enhancing the consistency and robustness of completion rate prediction in the continuous task phase. By introducing a completion state perception inference mechanism and inference attention information output, an explicit correlation is achieved between the prediction results and changes in load affordability and spatial accessibility. Based on this, an evidence summary and consistency correction mechanism are constructed to adaptively correct prediction deviations without changing the actual task behavior.

[0069] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A machine learning-based method for predicting the completion rate of sanitation vehicle operations, characterized in that: include, The current operating status of sanitation vehicles is hierarchically mapped to obtain operating status tokens. The operating status tokens are then concatenated and time step codes are added to form an operating status token sequence. The operation status token sequence is input into the conditional sequence machine learning model, and the completion state perception reasoning mechanism is used to reason about the evolution trend of sanitation vehicles from operation state to operation completion state, and generate the operation completion rate prediction value of the current operation stage. The aggregated conditional sequence machine learning model focuses on the changes in the intensity of attention to the job status token sequence during the inference process, forming an evidence sequence. The evidence sequence is then compressed to obtain an evidence summary. By binding the evidence summary with the job completion rate prediction value, a completion rate prediction result package is obtained. Based on the completion rate prediction results, the inference process of the conditional sequence machine learning model is corrected for consistency, and the corrected job completion rate prediction value is output.

2. The machine learning-based method for predicting the completion rate of sanitation vehicle operations as described in claim 1, characterized in that: The job status tokens include target completion rate instruction tokens, node compatibility awareness tokens, and visibility tokens.

3. The machine learning-based method for predicting the completion rate of sanitation vehicle operations as described in claim 2, characterized in that: The concatenation of the job status tokens refers to prioritizing the target completion rate instruction token as a global inference constraint input according to the causal hierarchy of the predicted job completion rate, followed by nodes compatible with perception tokens and visibility tokens.

4. The machine learning-based method for predicting the completion rate of sanitation vehicle operations as described in claim 1, characterized in that: The specific steps for reasoning about the evolution trend of sanitation vehicles from the working state to the working completed state, and generating a predicted value of the work completion rate for the current working stage, are as follows. Using the job status token sequence as input, a conditional sequence machine learning model is constructed based on an autoregressive sequence modeling structure. The conditional sequence machine learning model is trained using the historical job status token sequence as training samples. Input the current job status token sequence into the trained conditional sequence machine learning model to infer the job status from the job completion status and output the predicted job completion rate of the sanitation vehicle.

5. The machine learning-based method for predicting the completion rate of sanitation vehicle operations as described in claim 4, characterized in that: The specific steps for constructing the conditional sequence machine learning model are as follows. In the autoregressive sequence modeling structure, a causal mask is set to limit the temporal correlation range of the job status token sequence, thereby solidifying the temporal dependency constraints of the autoregressive sequence modeling structure into the structural inference rules of the conditional sequence machine learning model. The autoregressive sequence modeling structure performs time-step computation on the job state token sequence under causal mask constraints to obtain the intermediate implicit representation; The intermediate implicit representation is used as the input to the completion state perception reasoning mechanism to evaluate the evolution trend from the task state to the task completion state and generate a completion confidence representation. The conditional sequence machine learning model transforms the completion confidence representation into the current job completion rate prediction at each time step, thus completing the construction of the conditional sequence machine learning model.

6. The machine learning-based method for predicting the completion rate of sanitation vehicle operations as described in claim 5, characterized in that: The evaluation of the evolution trend from the task state to the task completion state refers to performing linear transformation and nonlinear mapping on the intermediate implicit representation to output a completion confidence representation.

7. The machine learning-based method for predicting the completion rate of sanitation vehicle operations as described in claim 5, characterized in that: The completion confidence representation refers to a quantitative indicator of the degree of confidence that the current job status continues to reach the job completion state.

8. The method for predicting the completion rate of sanitation vehicle operations based on machine learning as described in claim 1, characterized in that: The specific steps for binding the evidence summary with the predicted task completion rate are as follows: Time-align the evidence summary with the corresponding predicted task completion rate; The predicted task completion rate is used as the main prediction result identifier, and the evidence summary is used as the explanatory descriptive information on which the predicted task completion rate is formed, thus establishing a correspondence between the predicted task completion rate and the evidence summary. For each job completion rate prediction, the evidence digests of each job status token are associated and bound according to the correspondence between the job completion rate prediction and the evidence digest, to obtain the completion rate prediction result package.

9. The machine learning-based method for predicting the completion rate of sanitation vehicle operations as described in claim 1, characterized in that: The consistency correction process involves the following steps. Compare the predicted task completion rate with the long-term evolutionary trend reflected in the evidence summary over continuous time steps to see if there is any deviation. When deviations exist, the intermediate implicit representations formed during the inference process of the conditional sequence machine learning model are reweighted based on the relative contribution ratio of the main influencing factor categories in the completion rate prediction result package to the predicted value of the task completion rate. Conversely, if there is no deviation, there is no need for consistency correction.

10. The machine learning-based method for predicting the completion rate of sanitation vehicle operations as described in claim 9, characterized in that: The reweighting process refers to scaling the internal attention weights of the corresponding major influencing factor categories component by component using the normalized value of the relative contribution ratio.

Images