Power supply life prediction method and model training method and system for high-voltage ground wire repair robot

Abstract

The application provides a high-voltage ground wire repair robot power supply life prediction method, a model training method and a system, belonging to the technical field of battery health management. The method comprises the following steps: obtaining historical multi-dimensional behavior data; combining a Pearson correlation coefficient analysis method and a mutual information entropy method, performing double-criterion feature screening on the historical multi-dimensional behavior data to obtain core features related to fuel cell attenuation; performing entropy weight quantification on the information entropy of the core features to obtain core weighted features; performing model training on a general prediction model to be trained according to the core weighted features to obtain a trained general prediction model; performing energy analysis according to the historical multi-dimensional behavior data to obtain energy modes corresponding to different operation behaviors; and performing adaptive training on the general prediction model by using a transfer learning strategy according to the historical multi-dimensional behavior data and the energy modes to obtain a power supply life prediction model corresponding to each operation behavior. The application can improve the battery life prediction accuracy and scene adaptability of the model.

Description

Technical Field

[0001] This application relates to the field of battery health management technology, and in particular to a method for predicting the power life of a high-voltage grounding repair robot, a model training method, and a system. Background Technology

[0002] High-voltage grounding wire repair robots are key equipment for power system operation and maintenance. Their operating environments are often high-altitude and complex outdoor scenarios, placing extremely high demands on the stability and endurance of the power system. Currently, the power systems of high-voltage grounding wire repair robots generally use fuel cells as the core power supply component. This type of power supply has advantages such as high energy density and being environmentally friendly and pollution-free, meeting the robot's needs for long-term outdoor operations.

[0003] However, current control of power systems for high-voltage grounding repair robots mostly employs fixed-parameter control strategies, meaning that fixed operating parameters for the fuel cell are set based on preset work scenarios and load requirements. While some improved solutions introduce simple sensor data acquisition and feedback control, they can only adjust basic parameters in response to real-time load changes, without considering the impact of fuel cell system degradation characteristics and remaining lifespan on operating parameters.

[0004] Due to insufficient data utilization, low lifespan prediction accuracy, and poor adaptability of control strategies, the power system lifespan of high-voltage ground wire repair robots is shortened, maintenance costs are increased, and there is even a risk of equipment shutdown and work interruption caused by power system failure during high-altitude operations, which seriously affects the operation and maintenance efficiency of the power system. Summary of the Invention

[0005] The main objective of this application is to propose a power life prediction method, model training method, and system for a high-voltage grounding repair robot, aiming to improve the accuracy of battery life prediction and scenario adaptability of the model.

[0006] To achieve the above objectives, one aspect of this application proposes a training method for a power supply life prediction model of a high-voltage grounding wire repair robot, the training method comprising: Acquire historical multidimensional behavioral data of the high-voltage grounding wire repair robot, wherein the historical multidimensional behavioral data includes overall machine energy data, robot behavior data, fuel cell operation data, and environmental status data; By combining Pearson correlation coefficient analysis and mutual information entropy method, the historical multidimensional behavioral data is subjected to dual-criteria feature screening to obtain the core features related to fuel cell degradation. Entropy weight quantization is performed based on the information entropy of the core features to obtain the core weighted features; The general prediction model to be trained is trained according to the core weighted features to obtain a trained general prediction model. The general prediction model is obtained by introducing a multi-head attention mechanism on the basis of the time series prediction model. Energy analysis is performed based on the historical multidimensional behavioral data to obtain energy patterns corresponding to different operational behaviors, wherein the energy patterns include energy demand patterns and energy consumption patterns. Based on the historical multidimensional behavior data and the energy pattern, a transfer learning strategy is used to adapt and train the general prediction model to obtain a power life prediction model corresponding to each type of operation behavior.

[0007] In some embodiments, acquiring historical multidimensional behavioral data of the high-voltage grounding wire repair robot includes the following steps: The power output voltage, output current, output power, energy loss rate, and battery pack SOC of the high-voltage ground wire repair robot are collected by energy sensors to obtain the overall energy data of the robot. The robot's behavior data is obtained by collecting data on its speed, posture, load weight, and operation time using behavioral state sensors. The fuel cell operating data is obtained by collecting data such as stack temperature, stack humidity, hydrogen flow rate, air flow rate, anode and cathode pressure, single cell voltage consistency, and coolant flow rate from the high-voltage grounding repair robot using fuel cell status sensors. The environmental sensors collect data on the ambient temperature, humidity, wind speed, light intensity, and altitude of the high-voltage grounding wire repair robot to obtain environmental status data.

[0008] In some embodiments, the step of combining Pearson correlation coefficient analysis and mutual information entropy method to perform dual-criteria feature screening on the historical multidimensional behavioral data to obtain core features related to fuel cell degradation includes the following steps: Based on the historical multidimensional behavioral data and the preset fuel cell degradation index, the Pearson correlation coefficient is calculated to obtain the first correlation degree; The second correlation degree is obtained by calculating the mutual information entropy based on the historical multidimensional behavioral data and the fuel cell degradation index. The historical multidimensional behavioral data is filtered for features based on a preset first relevance threshold and the first relevance to obtain first relevance features. The historical multidimensional behavioral data is then filtered for features based on a preset second relevance threshold and the second relevance to obtain second relevance features. The core feature is obtained by the union of the first relevant feature and the second relevant feature.

[0009] In some embodiments, training the general prediction model to be trained based on the core weighting features to obtain a trained general prediction model includes the following steps: The core weighted features are divided according to time steps based on the temporal relationship to obtain time-series data; The time series data is modeled by performing feature temporal correlation modeling in parallel using multiple attention heads of different scales, and the temporal features output by each attention head are spliced ​​together to obtain multi-scale temporal features. Based on the multi-scale temporal features and static temporal features, temporal fusion is performed to obtain fused temporal features; Based on the fused temporal features, long-term dependencies are extracted and nonlinear transformations are performed to obtain dimensionality-reduced features; The dimensionality reduction features are input into the general prediction model to be trained to predict the power lifetime and obtain the prediction results, wherein the prediction results include the degradation level of the fuel cell system and the remaining life of the stack. The mean squared error is used as the loss function, and the parameters of the general prediction model are optimized based on the prediction results until the loss value of the prediction results is less than a preset performance threshold, thus obtaining a trained general prediction model.

[0010] In some embodiments, the step of performing energy analysis based on the historical multidimensional behavioral data to obtain energy patterns corresponding to different operational behaviors includes the following steps: Based on the overall machine energy data and the robot behavior data, energy demand clustering analysis is performed to obtain energy demand patterns corresponding to different operational behaviors; Energy consumption clustering analysis is performed based on the overall machine energy data and the robot behavior data to identify energy consumption patterns corresponding to different operational behaviors.

[0011] In some embodiments, the step of adapting and training the general prediction model using a transfer learning strategy based on the historical multidimensional behavioral data and the energy pattern to obtain a power lifetime prediction model corresponding to each operational behavior includes the following steps: Based on the influence weights of different environments on the energy consumption pattern, an attention mechanism is used to enhance the features of the environmental state data to obtain environmental enhancement features. The environmental enhancement features are weighted and fused with the overall machine energy data and the robot behavior data to obtain the fused features; The fused features of different work behaviors are input into a general prediction model for adaptation training using a transfer learning strategy to obtain a power life prediction model corresponding to each work behavior.

[0012] To achieve the above objectives, another aspect of this application proposes a method for predicting the power supply life of a high-voltage grounding wire repair robot, the prediction method comprising the following steps: Obtain the core behavioral data at the current moment, wherein the core behavioral data are the core features related to fuel cell degradation; Based on the core behavioral data, behavior matching is performed to obtain the current behavior pattern of the high-voltage grounding wire repair robot's current operation behavior; The pre-trained power lifetime prediction model is invoked based on the current behavior pattern, and the core behavior data is input into the power lifetime prediction model to predict the power lifetime and obtain the current prediction result. The power supply life prediction model is determined based on the aforementioned training method for the power supply life prediction model of the high-voltage grounding wire repair robot.

[0013] In some embodiments, the prediction method further includes the following steps: Based on the current prediction results, with the optimization objectives of minimizing the decay rate and maximizing the remaining lifetime, and with the operational energy requirements of the high-voltage ground wire repair robot as the constraint, the target operating parameters are controlled and optimized using the particle swarm optimization algorithm to obtain the optimized control strategy. The target operating parameters include hydrogen flow rate, air flow rate, stack cooling temperature, and output voltage. The control strategy is converted into a standard control signal, and the target operating parameters are adjusted according to the standard control signal.

[0014] To achieve the above objectives, another aspect of this application proposes a power supply life prediction model training system for a high-voltage grounding wire repair robot, the training system comprising: The first acquisition module is used to acquire historical multidimensional behavior data of the high-voltage grounding wire repair robot, wherein the historical multidimensional behavior data includes overall energy data, robot behavior data, fuel cell operation data and environmental status data; The feature filtering module is used to combine Pearson correlation coefficient analysis and mutual information entropy method to perform dual-criteria feature filtering on the historical multidimensional behavioral data to obtain the core features related to fuel cell degradation. The entropy weight quantization module is used to perform entropy weight quantization based on the information entropy of the core features to obtain the core weighted features. The pre-training module is used to train the general prediction model to be trained based on the core weighted features to obtain a trained general prediction model. The general prediction model is obtained by introducing a multi-head attention mechanism on the basis of the time series prediction model. The energy analysis module is used to perform energy analysis based on the historical multidimensional behavior data to obtain energy patterns corresponding to different work behaviors, wherein the energy patterns include energy demand patterns and energy consumption patterns. The transfer training module is used to adapt and train the general prediction model based on the historical multidimensional behavior data and the energy pattern using a transfer learning strategy, so as to obtain the power life prediction model corresponding to each type of operation behavior.

[0015] To achieve the above objectives, another aspect of this application proposes a power supply life prediction system for a high-voltage grounding wire repair robot, characterized in that the prediction system includes: The second acquisition module is used to acquire the core behavioral data at the current moment, wherein the core behavioral data is the core features related to fuel cell degradation; The behavior matching module is used to perform behavior matching based on the core behavior data to obtain the current behavior pattern of the high-voltage grounding wire repair robot's current operation behavior; The lifespan prediction module is used to call a pre-trained power supply lifespan prediction model based on the current behavior pattern, and input the core behavior data into the power supply lifespan prediction model to predict the power supply lifespan and obtain the current prediction result.

[0016] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0017] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0018] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0019] The embodiments of this application include at least the following beneficial effects: This application provides a method, model training method, and system for predicting the power life of a high-voltage grounding wire repair robot. The method includes acquiring historical multi-dimensional behavioral data of the high-voltage grounding wire repair robot, wherein the historical multi-dimensional behavioral data includes overall machine energy data, robot behavior data, fuel cell operation data, and environmental state data; combining Pearson correlation coefficient analysis and mutual information entropy method, performing dual-criteria feature screening on the historical multi-dimensional behavioral data to obtain core features related to fuel cell degradation; performing entropy weight quantization based on the information entropy of the core features to obtain core weighted features; training a general prediction model to be trained based on the core weighted features to obtain a trained general prediction model, wherein the general prediction model is obtained by introducing a multi-head attention mechanism on top of a time series prediction model; performing energy analysis based on the historical multi-dimensional behavioral data to obtain energy patterns corresponding to different operational behaviors, wherein the energy patterns include energy demand patterns and energy consumption patterns; and using a transfer learning strategy to adapt and train the general prediction model based on the historical multi-dimensional behavioral data and energy patterns to obtain a power life prediction model corresponding to each operational behavior. This scheme combines dual-criteria feature selection and entropy weighting to enhance long-term dependency modeling and focus on core factors, thereby improving the prediction accuracy of fuel cell degradation and remaining lifespan. It also incorporates transfer learning to adapt the model to different scenarios and improve its adaptability. Attached Figure Description

[0020] Figure 1 This is a flowchart of the power supply life prediction model training method for the high-voltage grounding wire repair robot provided in the embodiments of this application; Figure 2 This is a schematic diagram of the model structure of the general prediction model provided in the embodiments of this application; Figure 3 This is a flowchart of the power supply life prediction method for high-voltage ground wire repair robot provided in the embodiments of this application; Figure 4 This is a flowchart of a high-voltage grounding wire repair robot power supply life prediction method provided in another embodiment of this application; Figure 5 This is a closed-loop control block diagram of the optimized control strategy provided in the embodiments of this application; Figure 6 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of systems and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0023] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0024] (1) Information entropy: Information entropy is a mathematical concept proposed by Shannon that characterizes the amount of information, the purity of a set, and the degree of uncertainty. The more complex the system, the more different kinds of situations there are, and the larger its information entropy is. If a system is simpler, the fewer kinds of situations there are, the smaller its information entropy is.

[0025] (2) Mutual information entropy: Mutual information entropy is a dimensionless statistic that measures the information that one random variable can provide about the change of another random variable. It is used to measure the correlation between two random variables.

[0026] (3) Temporal Fusion Transformer (TFT): The TFT model is an advanced deep learning model designed specifically for time series forecasting. It combines multiple mechanisms of neural networks to handle complex relationships in time series data, aiming to deal with uncertainty and multi-scale dependencies in time series.

[0027] Next, the relevant technologies involved in the embodiments of this application will be described.

[0028] In related technologies, the power supply system of high-voltage ground wire repair robots has several core defects: First, data utilization is insufficient. Related technologies only collect a small amount of load or voltage and current data, and do not systematically collect multi-dimensional data such as overall machine energy, robot behavior, fuel cell operation and environmental conditions, which cannot fully reflect the key factors affecting the lifespan of the power system.

[0029] Second, the accuracy of life prediction is low. There is a lack of accurate prediction models for the degradation status of fuel cell systems and the remaining life of the stack, which makes it impossible to predict the trend of system performance decline in advance, resulting in a lag in the adjustment of control strategies.

[0030] Third, the control strategy has poor adaptability. Fixed parameters or simple feedback control cannot match the needs of different robot behavior modes and environmental changes. This can easily lead to a mismatch between operating parameters and actual needs, accelerating fuel cell degradation and shortening the lifespan of the power system.

[0031] Fourth, the model lacks a continuous optimization mechanism. Once the control model is trained, it remains fixed and cannot adapt to dynamic factors such as the aging of fuel cell components and long-term changes in the operating environment, resulting in a significant decrease in prediction accuracy and control performance in the later stages.

[0032] The aforementioned defects directly lead to a shortened lifespan and increased maintenance costs for the power system of the high-voltage grounding repair robot. They may even cause equipment shutdowns and work interruptions during high-altitude operations due to power system failures, seriously affecting the efficiency of power system operation and maintenance.

[0033] In view of this, to address at least one of the aforementioned deficiencies, this application provides a method, model training method, and system for predicting the power life of a high-voltage grounding wire repair robot. The method includes acquiring historical multi-dimensional behavioral data of the high-voltage grounding wire repair robot, including overall machine energy data, robot behavior data, fuel cell operation data, and environmental state data; combining Pearson correlation coefficient analysis and mutual information entropy analysis to perform dual-criteria feature screening on the historical multi-dimensional behavioral data to obtain core features related to fuel cell degradation; performing entropy weight quantification based on the information entropy of the core features to obtain core weighted features; training a general prediction model to be trained based on the core weighted features to obtain a trained general prediction model, wherein the general prediction model is obtained by introducing a multi-head attention mechanism on top of a time series prediction model; performing energy analysis based on the historical multi-dimensional behavioral data to obtain energy patterns corresponding to different operational behaviors, wherein the energy patterns include energy demand patterns and energy consumption patterns; and using a transfer learning strategy to adapt and train the general prediction model based on the historical multi-dimensional behavioral data and energy patterns to obtain a power life prediction model corresponding to each operational behavior. This scheme combines dual-criteria feature selection and entropy weighting to enhance long-term dependency modeling and focus on core factors, thereby improving the prediction accuracy of fuel cell degradation and remaining lifespan. It also incorporates transfer learning to adapt the model to different scenarios and improve its adaptability.

[0034] The high-voltage grounding wire repair robot power life prediction method, model training method, and system provided in this application relate to the field of battery health management technology. The high-voltage grounding wire repair robot power life prediction method and model training method provided in this application can be applied to a terminal, a server, or software running on a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the high-voltage grounding wire repair robot power life prediction method and model training method, but is not limited to the above forms.

[0035] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0036] Figure 1 This is an optional flowchart of the high-voltage grounding wire repair robot power supply life prediction model training method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S106.

[0037] Step S101: Obtain historical multidimensional behavior data of the high-voltage grounding wire repair robot.

[0038] Step S102: Combining Pearson correlation coefficient analysis and mutual information entropy method, the historical multidimensional behavioral data is screened using dual criteria to obtain the core features related to fuel cell degradation.

[0039] Step S103: Perform entropy weight quantization based on the information entropy of the core features to obtain the core weighted features.

[0040] Step S104: Train the general prediction model to be trained based on the core weighted features to obtain the trained general prediction model.

[0041] Step S105: Perform energy analysis based on historical multidimensional behavior data to obtain the energy patterns corresponding to different operational behaviors.

[0042] Step S106: Based on historical multidimensional behavior data and energy patterns, a transfer learning strategy is used to adapt and train the general prediction model to obtain a power life prediction model corresponding to each type of operation behavior.

[0043] In this embodiment, a multi-sensor system mounted on the high-voltage grounding wire repair robot continuously acquires four types of core data under normal driving and operation behaviors of the high-voltage grounding wire repair robot to obtain historical multi-dimensional behavioral data, which provides data support for subsequent model construction. Among them, the historical multi-dimensional behavioral data includes whole machine energy data reflecting the energy state of the high-voltage grounding wire repair robot itself, robot behavior data describing the operation actions of the high-voltage grounding wire repair robot, fuel cell operation data characterizing the internal working conditions of the fuel cell, and environmental state data recording the robot's operating environment conditions.

[0044] Optionally, to ensure data quality and consistency, systematic preprocessing of the collected original historical multidimensional behavioral data is required. This preprocessing includes deduplication, completion, anomaly removal, and normalization.

[0045] For example, a hash table deduplication algorithm is first used to remove duplicate data items (such as ambient temperature data that are completely identical in multiple consecutive frames). The collected multidimensional data samples are stored in the hash table as keys, and duplicate keys are automatically filtered out to retain unique and valid data.

[0046] For missing data caused by sensor failure or signal loss, linear interpolation based on time series is used to fill in missing data with a duration of less than 1 second, while the mean of adjacent time periods is used to fill in continuous missing data with a duration of more than 1 second.

[0047] Abnormal data is identified based on the 3σ principle, such as voltage or temperature data that are far beyond the normal range due to sensor fluctuations. Samples that exceed the normal value range are then removed.

[0048] The min-max normalization algorithm is used to map data of different dimensions and scales to the same scale, thus avoiding the negative impact of scale differences on model training performance.

[0049] After obtaining historical multidimensional behavioral data, in order to achieve accurate prediction of the degradation status of fuel cell system and the remaining life of stack, the correlation of the data was analyzed before model construction. Pearson correlation coefficient analysis and mutual information entropy method were used to screen the historical multidimensional behavioral data using dual criteria to identify the core features that are highly correlated with the degradation of fuel cell system.

[0050] By integrating two batches of features selected based on different analysis methods, the correlation between features and fuel cell degradation is measured from different perspectives, reducing the problem of reduced effective features caused by single-angle selection. Compared with feature selection using a single method, the dual-criteria feature selection in this embodiment can capture more comprehensive effective features.

[0051] Furthermore, the entropy weighting method is used to assign weights to the selected core features. After normalization, the information entropy of each core feature is calculated, and the weight is determined based on the magnitude of the information entropy to focus on key influencing factors. The core features are then weighted according to the weights determined by the information entropy to obtain the core weighted features. e i and weight w i The calculation formula is as follows: ; ; in, n The number of samples with core features. P ik For the first i The first feature k The normalized probability of each sample, where m is the total number of core features.

[0052] To enhance long-term dependency modeling and focus on core factors, this embodiment introduces a multi-head self-attention (MSA) mechanism on top of the time series prediction model during model construction. This results in an MSA-TFT model that integrates the multi-head self-attention mechanism, which serves as a general prediction model. Based on the core weighted features, the model learns the fuel cell degradation patterns and remaining lifespan trends of the high-voltage grounding repair robot, thereby improving the prediction accuracy of fuel cell degradation and remaining lifespan.

[0053] like Figure 2 As shown, the MSA-TFT deep learning model structure includes an input layer, a multi-head attention layer, a temporal fusion layer, a gated recurrent unit (GRU) layer, a fully connected layer, and an output layer. The data containing the core weighted features is divided into a training set and a validation set according to a certain ratio. The training set is used for iterative updates of the model parameters, and the validation set is used to evaluate the model performance.

[0054] The MSA-TFT model receives weighted core features through the input layer and then captures the temporal correlation of features through a multi-head attention mechanism to enhance the model's ability to model data with long-term dependencies.

[0055] The temporal fusion layer receives the output of the multi-head attention layer and uses feature mapping and joint learning to make the model focus on the feature types that have a significant impact.

[0056] The GRU layer further extracts long-term dependencies from the features to alleviate the gradient vanishing problem. Then, the output is sent to the fully connected layer to use an activation function to enhance the non-linear fitting ability of the model. Finally, the output layer uses a linear activation function to output the prediction result.

[0057] Training can be stopped when the performance evaluation results of the validation set show that the model does not decrease for several consecutive cycles, and a well-trained general prediction model is obtained.

[0058] To optimize the adaptability of the general prediction model to special scenarios, a clustering algorithm is used to perform energy analysis on the historical multidimensional behavior data of the high-voltage grounding wire repair robot for different operational behaviors (such as crossing obstacles, repair work, long-term continuous work, etc.). This automatically identifies the energy patterns corresponding to different specific operational behaviors, including energy demand patterns and energy consumption patterns.

[0059] Finally, based on the energy patterns obtained from the analysis, a transfer learning strategy was employed to specifically adapt and train the pre-trained general prediction model. The core idea of ​​transfer learning is to transfer the knowledge learned by the general model on a large amount of conventional data to the training of the model under specific operational scenarios. For each operational behavior, the multi-head attention layer and fully connected layer in the general prediction model were fine-tuned using its corresponding energy patterns and historical multi-dimensional behavioral data. This allows the general prediction model to better adapt to the power decay patterns under different operational behaviors while retaining general knowledge. After adaptation training, several power lifetime prediction models corresponding to different operational behaviors can be obtained. These power lifetime prediction models can be automatically matched and called according to the different operational behaviors of the robot during subsequent real-time prediction, thereby improving the battery prediction accuracy and scenario adaptability of the model.

[0060] In some embodiments, step S101 may include, but is not limited to, steps S201 to S204.

[0061] Step S201: Collect the power output voltage, output current, output power, energy loss rate, and battery pack SOC of the high-voltage grounding repair robot through the energy sensor to obtain the overall energy data of the robot.

[0062] Step S202: Collect the driving speed, working posture, load weight and working time of the high-voltage grounding repair robot through the behavior status sensor to obtain robot behavior data.

[0063] Step S203: Collect fuel cell operating data by using fuel cell status sensors to collect data on the stack temperature, stack humidity, hydrogen flow rate, air flow rate, anode and cathode pressure, single cell voltage consistency, and coolant flow rate of the high-voltage grounding repair robot.

[0064] Step S204: Collect environmental temperature, humidity, wind speed, light intensity and altitude of the high-voltage grounding wire repair robot through environmental sensors to obtain environmental status data.

[0065] In this embodiment, a multi-sensor system is built into the high-voltage grounding wire repair robot, including an energy sensor, a behavior status sensor, a fuel cell status sensor, and an environmental sensor, which can collect four types of core data in real time during the robot's normal driving and operation.

[0066] Specifically, the energy sensor is used to collect the overall energy data of the device, which includes data such as power supply output voltage, output current, output power, energy loss rate, and battery pack SOC (State of Charge), and the acquisition frequency is 10Hz.

[0067] Behavioral status sensors are used to collect robot behavior data, which includes driving speed, working posture (such as arm swing angle and working range), load weight, working time, etc., and the acquisition frequency is 5Hz.

[0068] Fuel cell status sensors are used to collect fuel cell operating data, including stack temperature, stack humidity, hydrogen flow rate, air flow rate, anode and cathode pressure, single cell voltage consistency, coolant flow rate, etc., with a collection frequency of 10Hz.

[0069] Environmental sensors are used to collect environmental status data, including ambient temperature, humidity, wind speed, light intensity, and altitude. These data are collected by the environmental sensors mounted on the robot at a frequency of 2Hz.

[0070] In some embodiments, step S102 may include, but is not limited to, steps S301 to S304.

[0071] Step S301: Calculate the Pearson correlation coefficient based on historical multidimensional behavioral data and preset fuel cell degradation index to obtain the first correlation degree.

[0072] Step S302: Calculate the mutual information entropy based on historical multidimensional behavioral data and fuel cell degradation index to obtain the second correlation degree.

[0073] Step S303: Perform feature filtering on historical multidimensional behavioral data according to a preset first correlation threshold and a first correlation degree to obtain first correlation features, and perform feature filtering on historical multidimensional behavioral data according to a preset second correlation threshold and a second correlation degree to obtain second correlation features.

[0074] Step S304: Obtain the core feature based on the union of the first relevant feature and the second relevant feature.

[0075] In this embodiment, the preset fuel cell degradation index is a parameter used to quantify the health status of the fuel cell, including but not limited to indices such as stack capacity degradation rate and single cell voltage degradation magnitude. Different fuel cell degradation indices belong to different dimensions. The preset first correlation threshold is a threshold for measuring whether data is related to the fuel cell degradation index under the Pearson correlation coefficient method, and the preset second correlation threshold is a threshold for measuring whether data is related to the fuel cell degradation index under the mutual information entropy method.

[0076] Specifically, in this embodiment, the Pearson coefficient analysis method and the mutual information entropy method are used to perform correlation analysis on each type of data in the historical multidimensional behavioral data with each preset fuel cell attenuation index. When the historical multidimensional behavioral data satisfies that it is correlated with at least one of the fuel cell attenuation in either of the two analysis methods, the data can be confirmed as the core feature.

[0077] For example, assuming that both the first and second correlation thresholds are 0.6, the fuel cell degradation indicators include the stack capacity degradation rate and the single cell voltage degradation magnitude. As long as the absolute value of the correlation coefficient of one of the degradation indicators is ≥0.6, the feature can be retained.

[0078] The Pearson correlation coefficients between each data feature in the historical multidimensional behavioral data and the fuel cell degradation index are calculated to obtain the first correlation degree. Based on the first correlation degree and a preset first correlation threshold, feature selection is performed, retaining features with an absolute Pearson correlation coefficient ≥ 0.6 for any fuel cell degradation index, thus obtaining the first correlation features. The Pearson correlation coefficient calculation method is as follows: ; r The Pearson correlation coefficient, also known as the first correlation, has a range of [-1, 1]. The closer the absolute value is to 1, the stronger the correlation between the two variables. The feature that satisfies the condition of absolute value ≥ 0.6 is retained. x k For the first k One data feature sample; y kFor the first k A sample of fuel cell degradation indicators, μ x Let x be the mean of the data feature. μ y Sample of fuel cell degradation index y The mean, p This represents the number of samples.

[0079] Simultaneously, the mutual information entropy between each feature and the fuel cell degradation index is calculated to obtain the second correlation degree. Based on the second correlation degree and a preset second correlation threshold, feature selection is performed, retaining features with a mutual information entropy ≥ 0.6 for any fuel cell degradation index, thus obtaining the second correlation features. The mutual information entropy calculation method is as follows: ; in, I ( X ; Y ) is a random variable X (Data characteristics) and Y Mutual information entropy (of fuel cell degradation indicators); P ( x ) represents the marginal probability of feature x. P ( y ( ) is the fuel cell degradation index y The marginal probability, P ( x , y (x) represents the characteristic x and the fuel cell degradation index y The joint probability.

[0080] Finally, the two screening results are merged, and the core features are obtained by integrating the union of the first and second relevant features, making the captured effective features more comprehensive.

[0081] In some embodiments, step S104 may include, but is not limited to, steps S401 to S406.

[0082] Step S401: Divide the core weighted features according to the time step based on the temporal relationship to obtain time series data.

[0083] Step S402 involves using multiple attention heads of different scales to perform feature temporal correlation modeling on the temporal data in parallel, and then concatenating the temporal features output by each attention head to obtain multi-scale temporal features.

[0084] Step S403: Perform time fusion based on multi-scale time series features and static time series features to obtain fused time series features.

[0085] Step S404: Extract long-term dependencies and perform nonlinear transformation based on the fused temporal features to obtain dimensionality-reduced features.

[0086] Step S405: Input the dimensionality reduction features into the general prediction model to be trained to predict power lifetime and obtain the prediction results.

[0087] Step S406: Using mean squared error as the loss function, the parameters of the general prediction model are optimized based on the prediction results until the loss value of the prediction results is less than the preset performance threshold, thus obtaining the trained general prediction model.

[0088] In this embodiment, the MSA-TFT model is specifically designed as follows: The input layer receives weighted core features. The input dimension is the number of core weighted features × the time step, and the time step is set to 60s, that is, the input is time series data within 60s.

[0089] The multi-head attention layer uses eight attention heads, each of which independently computes the temporal correlations of features in the temporal data provided by the input layer. By concatenating the outputs of different attention heads, multi-scale temporal features are captured, enhancing the model's ability to model long-term dependent data, as shown below: ; ; in, h To focus on the number of heads, Q (Query matrix) K (Key matrix) V The (value matrix) is obtained by linear transformation of the input layer data, and all three dimensions are (time step × feature dimension). d k for Q and K The dimension is used to scale the attention score and avoid excessively large values; head i For the first i The output of each attention head; W o This is a weighted matrix for multi-head output.

[0090] The temporal fusion layer first performs temporal dimension expansion and linear mapping on the static features to achieve dimensional alignment with the dynamic temporal features. Static features are added to the model as independent feature dimensions during the temporal fusion layer stage. Static features do not change over time and are usually basic information inherent to the device, such as fuel cell model, production date, rated power, etc. Dynamic temporal features are the multi-scale temporal features output by the multi-head attention layer. These features are deep encoding features of dynamic temporal features and only contain the dynamic temporal feature dimension.

[0091] Then, the two types of features are jointly learned through a gating mechanism. The sigmoid activation function is used to output adaptive weights in the 0-1 interval and normalize them to dynamically adjust the fusion ratio of static and dynamic time-series features. Subsequently, the weights are fused with the corresponding features element by element to perform weighted processing and complete the fusion. Finally, the fused time-series features that integrate the inherent basic information of the equipment and the real-time dynamic information of the operation are output to the next layer, so that the model focuses on the feature types that have a significant impact.

[0092] The GRU layer and fully connected layer are used for long-term dependency extraction and non-linear fitting, respectively. The GRU layer has two GRU units, each with a hidden layer dimension of 256. The fully connected layer has two fully connected layers, with the first layer having an output dimension of 128 and the second layer having an output dimension of 64. The ReLU activation function is used to enhance the model's non-linear fitting capability. The ReLU activation function formula is: ReLU ( x ) = max(0, x ).

[0093] The output layer uses a linear activation function to output two prediction results: the degradation status of the fuel cell system (represented by degradation levels 0-5, where level 0 indicates no degradation and level 5 indicates severe degradation) and the remaining lifespan of the fuel cell stack (represented by the number of remaining operating hours).

[0094] During model training, mean squared error (MSE) is used as the loss function: ; in, q To assess sample size, y i For the first i The true value of each sample This is the model prediction value for the i-th sample.

[0095] The Adam optimizer was used for parameter optimization, with the initial learning rate set to 0.001. A learning rate decay strategy was employed (the learning rate decayed to 0.9 every 10 epochs), and the number of training epochs was set to 100. Training stopped when the validation set loss function value did not decrease for 5 consecutive epochs. The model performance was evaluated using mean absolute error (MAE) and root mean square error (RMSE), with requirements of MAE ≤ 0.05 (decay level prediction) and MAE ≤ 20h (remaining lifetime prediction) to ensure good generalization ability of the model.

[0096] In some embodiments, step S105 may include, but is not limited to, steps S501 to S502.

[0097] Step S501: Perform energy demand clustering analysis based on the overall machine energy data and robot behavior data to obtain energy demand patterns corresponding to different operational behaviors.

[0098] Step S502: Perform energy consumption cluster analysis based on the overall machine energy data and robot behavior data to identify energy consumption patterns corresponding to different operational behaviors.

[0099] In this embodiment, the robot control system presets specific work scenarios and collects the overall energy data (output power, energy loss rate, etc.) and robot behavior data (work posture, load weight, work speed, etc.) of the high-voltage grounding repair robot under different specific work behaviors. At least 100 sets of valid data samples are collected for each specific behavior.

[0100] K-means clustering algorithm is used to perform cluster analysis on the collected whole machine energy data and robot behavior data under specific operation behaviors to identify energy demand patterns (such as high load high energy demand mode, long-term low load continuous demand mode, etc.) and energy consumption patterns (such as rapid consumption mode, steady consumption mode) corresponding to different specific operation behaviors. Feature cluster centers of different behavior patterns are obtained. The specific calculation method is as follows: ; in, c j For the first j The center of each cluster represents the core feature of the corresponding energy pattern; S j For the first j The set of samples contained in each cluster; s j For the first j The number of samples in each cluster.

[0101] For example, the predefined energy demand modes in this embodiment include a high load high energy demand mode and a long-term low load continuous demand mode. The high load high energy demand mode is characterized by a load weight of not less than 70% of the rated load, an average output power of not less than 30W, and a demand duration of 5-30 minutes. The long-term low load continuous demand mode is characterized by a load weight of not more than 30% of the rated load, an average output power of not more than 10W, and a demand duration of not less than 60 minutes.

[0102] Energy consumption modes include rapid consumption mode, stable consumption mode, and intermittent consumption mode. Rapid consumption mode corresponds to high load and high energy demand mode, characterized by an energy loss rate of no less than 8% / h, a power output fluctuation rate of no more than 5%, and a fast energy consumption rate. Stable consumption mode corresponds to long-term low load and continuous demand mode, characterized by an energy loss rate of no more than 3% / h, a power output fluctuation rate of no more than 3%, and a stable energy consumption rate. Intermittent consumption mode corresponds to instantaneous high load and pulse demand mode, characterized by an energy loss rate of 4%-6% / h, a power output fluctuation rate of 10%-15%, and intermittent energy consumption.

[0103] Based on the output power in the overall energy data and the load weight and operation time in the robot behavior data, the energy demand pattern of the high-voltage grounding wire repair robot under different operation behaviors can be analyzed. At the same time, based on the energy loss rate and output power in the overall energy data and the operation time in the robot behavior data, the energy consumption pattern of the high-voltage grounding wire repair robot under different operation behaviors can be analyzed.

[0104] The K-means clustering algorithm utilizes the energy feature similarity of specific job behaviors to determine the number of cluster centers in an unsupervised manner based on the actual feature distribution of the data.

[0105] In some embodiments, step S106 may include, but is not limited to, steps S601 to S603.

[0106] Step S601: Based on the influence weights of different environments on energy consumption patterns, an attention mechanism is used to enhance the features of the environmental state data to obtain enhanced environmental features.

[0107] Step S602: The environmental enhancement features are weighted and fused with the whole machine energy data and robot behavior data to obtain the fused features.

[0108] Step S603: Using a transfer learning strategy, the fusion features of different work behaviors are input into the general prediction model for adaptation training to obtain the power life prediction model corresponding to each work behavior.

[0109] In this embodiment, by analyzing the impact of different environmental state data on the overall energy data and robot behavior data under specific operational behaviors, the influence weight of different environmental state data on energy consumption patterns is calculated, and an attention mechanism is used to enhance the features of the environmental state data to obtain environmental enhancement features.

[0110] It should be noted that when calculating the weights of the influence of different environmental conditions on energy consumption patterns, since the robot's energy demand pattern is the basic energy requirement for it to complete specific tasks, and environmental characteristics will change the robot's workload resistance, directly affecting the robot's actual energy demand, which in turn leads to a synchronous change in the energy consumption pattern of the fuel cell, the weight calculation of energy consumption patterns already includes the indirect influence of environmental characteristics on energy demand patterns.

[0111] Subsequently, using fuel cell operation data under specific operational behaviors as the core foundation, the weighted environmental enhancement is fused with the energy behavior features formed by the whole machine energy data and robot behavior data. After dimensional alignment and weighted fusion, a multi-dimensional fused feature set is formed, resulting in fused features, which improves the model's adaptability to dynamic environmental changes.

[0112] For example, let the core feature vector of energy consumption pattern be... Q (Dimension [1, d k ]), Environmental state data feature matrix K = V (dimension[) n , d k The influence weights of environmental characteristics on energy consumption patterns and environmental enhancement characteristics are calculated using the following formula: First, calculate the raw attention score: ; Scaling the scores helps prevent gradient explosion. ; Normalized influence weights are generated using the softmax activation function. α : ; Generate a weighted environment enhancement feature matrix V’ : ; Enhance the environmental feature matrix V’ With energy behavior characteristic matrix F Weighted summation yields the enhanced feature set of fuel cell operating data. Ffusion : ; In the above formula, d k For feature dimension, n For the quantity of environmental characteristics, λ The fusion coefficient (value range [0,1]) can be dynamically adjusted according to the robot's operating environment.

[0113] The fused features obtained by weighted fusion are input into the pre-trained general prediction model in step S104 for adaptation training. During training, a transfer learning strategy is adopted, using the model parameters of the general prediction model as initial parameters, and only fine-tuning the parameters of the multi-head attention layer and the fully connected layer to reduce training costs. The loss function and optimizer are the same as in step S406. Training continues until the prediction error on the validation set reaches a preset threshold, resulting in a power lifetime prediction model adapted to each specific operation behavior of the robot.

[0114] Figure 3 This is an optional flowchart of the high-voltage grounding wire repair robot power life prediction method provided in the embodiments of this application. Figure 3 The prediction method may include, but is not limited to, steps S701 to S703.

[0115] Step S701: Obtain the core behavioral data at the current moment.

[0116] Step S702: Perform behavior matching based on core behavior data to obtain the current behavior pattern of the high-voltage grounding wire repair robot's current operation behavior.

[0117] Step S703: Based on the current behavior pattern, call the pre-trained power lifetime prediction model, and input the core behavior data into the power lifetime prediction model to predict the power lifetime and obtain the current prediction result.

[0118] In this embodiment, the power life prediction method for the high-voltage grounding repair robot is used to achieve real-time monitoring and prediction of the fuel cell system status during robot operation.

[0119] Specifically, the sensor system on the robot collects core features that are highly correlated with the degradation of the fuel cell system in real time through step S102, and obtains core behavioral data covering four types of data: current operating data of the fuel cell, energy data of the whole machine, robot behavior data, and environmental status data.

[0120] The real-time data is then deduplicated, completed, anomaly removed, and normalized through a preprocessing workflow, with a processing delay of ≤0.5s.

[0121] Next, by matching the core behavioral data with the cluster centers determined by the clustering algorithm, the current behavior pattern (such as routine behavior or specific behavior) of the high-voltage grounding repair robot is determined. Based on the current behavior pattern, the pre-trained power life prediction model is called, and the pre-processed real-time data is input into the model. The model outputs the current fuel cell system degradation level and stack remaining life in real time, with a prediction delay of ≤1s, to ensure the timeliness of the prediction results.

[0122] It is understandable that the power supply life prediction model is determined based on the aforementioned training method for the power supply life prediction model of the high-voltage grounding wire repair robot.

[0123] In some embodiments, the power life prediction method for high-voltage grounding repair robots may also include, but is not limited to, steps S801 to S802.

[0124] Step S801: Based on the current prediction results, with the optimization objective of minimizing the decay rate and maximizing the remaining lifespan, and with the operational energy requirements of the high-voltage ground wire repair robot as the constraint, the target operating parameters are controlled and optimized using the particle swarm optimization algorithm to obtain the optimized control strategy.

[0125] Step S802: The control strategy is converted into a standard control signal, and the target operating parameters are adjusted according to the standard control signal.

[0126] In this embodiment, the target operating parameters of the fuel cell are dynamically optimized based on the current prediction results in real time, thereby delaying system degradation.

[0127] Specifically, based on the current prediction results in step S703, an improved particle swarm optimization (PSO) algorithm is used to generate an optimized control strategy for the fuel cell system. The optimization objective is to minimize the decay rate and maximize the remaining lifetime, with the constraint being the energy demand for robot operation. The target operating parameters for optimization include hydrogen flow rate (adjustment range: 0.1-0.5 L / min), air flow rate (adjustment range: 0.5-2.0 L / min), stack cooling temperature (adjustment range: 60-80℃), and output voltage (adjustment range: 24-36V). The optimal combination of operating parameters and their adjustment range are obtained through iterative solving using the particle swarm optimization algorithm, calculated as follows: ; ; The optimization objective of the algorithm is to minimize the fuel cell degradation rate and maximize the remaining stack lifetime. x Target running parameters; v i，t For the first i Particles t The speed of time; x i，t For the first i Particles t The position at any given time corresponds to the current value of the target's operating parameters; ω Inertial weights are used to balance the algorithm's global search capability with its local search capability. c 1. c 2 is the learning factor, which adjusts the weights of particles moving closer to their own historical best position and the global best position of the population, respectively.r 1. r 2 is a random number that follows a uniform distribution in [0,1].

[0128] The optimized control strategy is converted into standardized control signals and transmitted to the actuators of the fuel cell system via the CAN bus. The actuators adjust their operating parameters in real time according to the control signals. During the adjustment process, the adjusted operating data is collected in real time and fed back to the MSA-TFT model to form a closed-loop control, ensuring that the adjustment effect meets the optimization target.

[0129] The following is a detailed description and explanation of the solutions in the embodiments of the present invention, using specific application examples: Reference Figure 4 The embodiments of this application are based on a closed-loop design of the entire process of "data acquisition - model building - pattern adaptation - real-time prediction - optimization control - model update", providing a full-link closed-loop system to achieve coordination from data processing to control output to model iteration.

[0130] Specifically, during the model building and training phase, multidimensional data of conventional robot behavior are collected and preprocessed to obtain historical multidimensional behavior data, including overall robot energy data, robot behavior data, fuel cell operation data, and environmental state data.

[0131] To ensure data consistency, systematic preprocessing was performed on the collected raw data.

[0132] First, data items that are completely identical across multiple consecutive frames due to repeated acquisition are removed.

[0133] Then, for missing data, if the missing duration is less than 1 second, linear interpolation based on time series is used for completion; if the missing duration is more than 1 second, the mean of adjacent time periods is used for filling. The calculation methods for linear interpolation and mean filling are as follows: ; ; in, xt for t The missing data values ​​at any given time, i.e. the target data to be completed; x t-1 for t The closest valid data value before the current time. t t-1 Its corresponding timestamp; x t+1 for t The closest valid data value after time step [time] t t+1 Its corresponding timestamp; t This is the timestamp corresponding to the missing data.

[0134] Based on the 3σ principle, identify abnormal data and remove those exceeding […]. μ -3σ, μ Samples within the range of +3σ were removed, as shown below: ; in, x The original data to be judged; μ σ is the mean of the corresponding data dimension, i.e., the average of all valid original data in that dimension; σ is the standard deviation of the corresponding data dimension, used to measure the dispersion of the data; μ -3σ, μ [+3σ] represents the normal data range, which covers 99.73% of normal data. Data outside this range will be considered abnormal and removed.

[0135] Finally, the min-max normalization method is used to eliminate the dimensional differences between data of different dimensions: ; in, x This refers to the raw data, i.e., a single sample from the collected multidimensional data; x’ For the normalized standard data, max( x ) represents the maximum value in the corresponding dimension of the original data, min( x () represents the minimum value in the corresponding original data dimension.

[0136] Based on preprocessed historical multidimensional behavioral data, an improved TFT (Temporal Fusion Transformer) model (MSA-TFT model) integrating a multi-head attention mechanism is constructed to achieve accurate prediction of fuel cell system degradation and stack remaining life.

[0137] Furthermore, by utilizing multidimensional behavioral data of specific robots, energy patterns are analyzed and the model is trained through transfer learning to obtain a power lifetime prediction model adapted to specific behavioral patterns.

[0138] During the real-time prediction phase, current operating data is collected in real time, and behavior patterns are matched with the operating data to determine whether the robot's current behavior pattern is routine or specific. Based on the identified behavior pattern, the corresponding MSA-TFT model is called to monitor and predict the state of the fuel cell system in real time during the robot's operation, thereby obtaining the current fuel cell system degradation level and the remaining life of the stack.

[0139] Based on the real-time prediction results, an optimized control strategy is generated, and the operating parameters are adjusted in real time. The model is updated and trained by continuously collecting operating data after parameter adjustment and execution.

[0140] like Figure 5 As shown, during robot operation, data on fuel cell system operation, overall energy, robot behavior, and environment are continuously collected to establish a historical database. Two update trigger conditions are set: one is time-based (every 7 days), and the other is data volume-based (when 500 new data sets are added). Meeting either condition initiates the model update process.

[0141] The newly added data is merged with the data in the historical database and processed according to the preprocessing procedure, and the training set (70%), validation set (20%), and test set (10%) are re-divided. The MSA-TFT model is retrained using the model structure and training parameters from step S104. After training, the model performance is verified using the test set. If the test set performance meets the requirements, the new model is deployed to replace the original model; if it does not meet the requirements, the training parameters (such as learning rate and number of epochs) are readjusted and the model is trained again until the performance meets the requirements. Through model updates, dynamic factors such as the aging of fuel cell components and long-term changes in the operating environment are adapted to ensure the long-term effectiveness of the model.

[0142] In summary, the end-to-end closed-loop system constructed in this application embodiment enables coordination from data processing to control output and model iteration. The MSA-TFT model, which integrates a multi-head attention mechanism, combined with dual-criteria feature selection and entropy weighting, strengthens long-term dependency modeling and core factor focus, thus improving the prediction accuracy of fuel cell degradation and remaining lifespan. Furthermore, this application embodiment identifies specific behavioral energy patterns through K-means clustering and achieves model scenario adaptation through transfer learning; it uses an improved PSO algorithm to dynamically optimize key operating parameters, coupled with closed-loop feedback control, to slow down system degradation.

[0143] Compared with related technologies, the embodiments of this application have several significant beneficial effects: (1) High prediction accuracy: By constructing an MSA-TFT model that integrates multi-head attention mechanism, combined with multi-dimensional data correlation analysis and weight assignment, the system can accurately predict the degradation status of fuel cell system and the remaining life of stack.

[0144] (2) It is highly adaptable. Through the hierarchical collection of data on routine and specific behaviors and model adaptation training, combined with environmental feature enhancement, the model can accurately match different working modes of the robot and complex environmental changes.

[0145] (3) Significantly extended lifespan: Based on the dynamic optimization control strategy of real-time prediction results, the operating parameters of the fuel cell can be precisely adjusted, effectively delaying system degradation.

[0146] (4) It has good long-term effectiveness. Through continuous data collection and model update mechanism, it adapts to component aging and environmental changes, ensures the long-term prediction accuracy and optimization effect of the model, reduces the later maintenance cost, and improves the reliability of robot operation.

[0147] (5) Excellent real-time performance. The delay of the entire data preprocessing, model prediction and control adjustment process is ≤2s, which can meet the real-time control requirements of robot high-altitude operation and avoid safety hazards caused by control lag.

[0148] This application also provides a training system for a power supply life prediction model of a high-voltage grounding wire repair robot, which can implement the above-mentioned training method. The training system includes: The first acquisition module is used to acquire historical multidimensional behavioral data of the high-voltage grounding wire repair robot. This historical multidimensional behavioral data includes overall energy data, robot behavior data, fuel cell operation data, and environmental status data.

[0149] The feature selection module combines Pearson correlation coefficient analysis and mutual information entropy method to perform dual-criteria feature selection on historical multidimensional behavioral data to obtain core features related to fuel cell degradation.

[0150] The entropy weight quantization module is used to perform entropy weight quantization based on the information entropy of the core features to obtain the core weighted features.

[0151] The pre-training module is used to train the general prediction model to be trained based on the core weighted features, so as to obtain the trained general prediction model. The general prediction model is obtained by introducing a multi-head attention mechanism on the basis of the time series prediction model.

[0152] The energy analysis module is used to perform energy analysis based on historical multidimensional behavioral data to obtain energy patterns corresponding to different operational behaviors. These energy patterns include energy demand patterns and energy consumption patterns.

[0153] The transfer training module is used to adapt and train a general prediction model based on historical multidimensional behavioral data and energy patterns using a transfer learning strategy, so as to obtain a power life prediction model corresponding to each type of operation behavior.

[0154] It is understood that the content of the above training method embodiments is applicable to this training system embodiment. The specific functions implemented by this training system embodiment are the same as those of the above training method embodiments, and the beneficial effects achieved are also the same as those achieved by the above training method embodiments.

[0155] This application also provides a power supply life prediction system for a high-voltage grounding wire repair robot, which can implement the above-mentioned prediction method. The prediction system includes: The second acquisition module is used to acquire the core behavioral data at the current moment, which are the core features related to fuel cell degradation.

[0156] The behavior matching module is used to perform behavior matching based on core behavior data to obtain the current behavior pattern of the high-voltage grounding wire repair robot's current operation behavior.

[0157] The lifespan prediction module is used to call a pre-trained power supply lifespan prediction model based on the current behavior pattern, and input the core behavior data into the power supply lifespan prediction model to predict the power supply lifespan and obtain the current prediction result.

[0158] It is understood that the content of the above prediction method embodiments is applicable to this prediction system embodiment. The specific functions implemented by this prediction system embodiment are the same as those of the above prediction method embodiments, and the beneficial effects achieved are also the same as those achieved by the above prediction method embodiments.

[0159] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described high-voltage grounding wire repair robot power life prediction model training method or high-voltage grounding wire repair robot power life prediction method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0160] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0161] Reference Figure 6 , Figure 6 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0162] The memory 902 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 902 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and is called and executed by the processor 901 to execute the high-voltage grounding wire repair robot power life prediction model training method or the high-voltage grounding wire repair robot power life prediction method described above in the embodiments of this application.

[0163] The input / output interface 903 is used to implement information input and output.

[0164] The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0165] Bus 905 transmits information between various components of the device, such as processor 901, memory 902, input / output interface 903, and communication interface 904.

[0166] The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0167] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described high-voltage ground wire repair robot power life prediction model training method or high-voltage ground wire repair robot power life prediction method.

[0168] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0169] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described high-voltage ground wire repair robot power life prediction model training method or high-voltage ground wire repair robot power life prediction method.

[0170] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0171] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0172] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0173] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0174] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0175] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0176] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0177] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0178] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0179] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0180] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0181] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0182] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A training method for a power supply life prediction model of a high-voltage grounding wire repair robot, characterized in that, The training method includes the following steps: Acquire historical multidimensional behavioral data of the high-voltage grounding wire repair robot, wherein the historical multidimensional behavioral data includes overall machine energy data, robot behavior data, fuel cell operation data, and environmental status data; By combining Pearson correlation coefficient analysis and mutual information entropy method, the historical multidimensional behavioral data is subjected to dual-criteria feature screening to obtain the core features related to fuel cell degradation. Entropy weight quantization is performed based on the information entropy of the core features to obtain the core weighted features; The general prediction model to be trained is trained according to the core weighted features to obtain a trained general prediction model. The general prediction model is obtained by introducing a multi-head attention mechanism on the basis of the time series prediction model. Energy analysis is performed based on the historical multidimensional behavioral data to obtain energy patterns corresponding to different operational behaviors, wherein the energy patterns include energy demand patterns and energy consumption patterns. Based on the historical multidimensional behavior data and the energy pattern, a transfer learning strategy is used to adapt and train the general prediction model to obtain a power life prediction model corresponding to each type of operation behavior.

2. The method according to claim 1, characterized in that, The acquisition of historical multidimensional behavioral data of the high-voltage grounding wire repair robot includes the following steps: The power output voltage, output current, output power, energy loss rate, and battery pack SOC of the high-voltage ground wire repair robot are collected by energy sensors to obtain the overall energy data of the robot. The robot's behavior data is obtained by collecting data on its speed, posture, load weight, and operation time using behavioral state sensors. The fuel cell operating data is obtained by collecting data such as stack temperature, stack humidity, hydrogen flow rate, air flow rate, anode and cathode pressure, single cell voltage consistency, and coolant flow rate from the high-voltage grounding repair robot using fuel cell status sensors. The environmental sensors collect data on the ambient temperature, humidity, wind speed, light intensity, and altitude of the high-voltage grounding wire repair robot to obtain environmental status data.

3. The method according to claim 1, characterized in that, The method of combining Pearson correlation coefficient analysis and mutual information entropy method to perform dual-criteria feature screening on the historical multidimensional behavioral data to obtain the core features related to fuel cell degradation includes the following steps: Based on the historical multidimensional behavioral data and the preset fuel cell degradation index, the Pearson correlation coefficient is calculated to obtain the first correlation degree; The second correlation degree is obtained by calculating the mutual information entropy based on the historical multidimensional behavioral data and the fuel cell degradation index. The historical multidimensional behavioral data is filtered for features based on a preset first relevance threshold and the first relevance to obtain first relevance features. The historical multidimensional behavioral data is then filtered for features based on a preset second relevance threshold and the second relevance to obtain second relevance features. The core feature is obtained by the union of the first relevant feature and the second relevant feature.

4. The method according to claim 1, characterized in that, The process of training the general prediction model to be trained based on the core weighted features to obtain the trained general prediction model includes the following steps: The core weighted features are divided according to time steps based on the temporal relationship to obtain time-series data; The time series data is modeled by performing feature temporal correlation modeling in parallel using multiple attention heads of different scales, and the temporal features output by each attention head are spliced ​​together to obtain multi-scale temporal features. Based on the multi-scale temporal features and static temporal features, temporal fusion is performed to obtain fused temporal features; Based on the fused temporal features, long-term dependencies are extracted and nonlinear transformations are performed to obtain dimensionality-reduced features; The dimensionality reduction features are input into the general prediction model to be trained to predict the power lifetime and obtain the prediction results, wherein the prediction results include the degradation level of the fuel cell system and the remaining life of the stack. The mean squared error is used as the loss function, and the parameters of the general prediction model are optimized based on the prediction results until the loss value of the prediction results is less than a preset performance threshold, thus obtaining a trained general prediction model.

5. The method according to claim 1, characterized in that, The step of performing energy analysis based on the historical multidimensional behavioral data to obtain energy patterns corresponding to different operational behaviors includes the following steps: Based on the overall machine energy data and the robot behavior data, energy demand clustering analysis is performed to obtain energy demand patterns corresponding to different operational behaviors; Energy consumption clustering analysis is performed based on the overall machine energy data and the robot behavior data to identify energy consumption patterns corresponding to different operational behaviors.

6. The method according to claim 5, characterized in that, The step of adapting and training the general prediction model using a transfer learning strategy based on the historical multidimensional behavior data and the energy pattern to obtain a power lifetime prediction model corresponding to each type of operation includes the following steps: Based on the influence weights of different environments on the energy consumption pattern, an attention mechanism is used to enhance the features of the environmental state data to obtain environmental enhancement features. The environmental enhancement features are weighted and fused with the overall machine energy data and the robot behavior data to obtain the fused features; The fused features of different work behaviors are input into a general prediction model for adaptation training using a transfer learning strategy to obtain a power life prediction model corresponding to each work behavior.

7. A method for predicting the power supply life of a high-voltage grounding wire repair robot, characterized in that, The prediction method includes the following steps: Obtain core behavioral data at the current moment, wherein the core behavioral data are core features related to fuel cell degradation; Based on the core behavioral data, behavior matching is performed to obtain the current behavior pattern of the high-voltage grounding wire repair robot's current operation behavior; The pre-trained power lifetime prediction model is invoked based on the current behavior pattern, and the core behavior data is input into the power lifetime prediction model to predict the power lifetime and obtain the current prediction result. The power supply life prediction model is determined by the training method for the power supply life prediction model of the high-voltage grounding wire repair robot according to any one of claims 1-6.

8. The method according to claim 7, characterized in that, The prediction method further includes the following steps: Based on the current prediction results, with the optimization objectives of minimizing the decay rate and maximizing the remaining lifetime, and with the operational energy requirements of the high-voltage ground wire repair robot as the constraint, the target operating parameters are controlled and optimized using the particle swarm optimization algorithm to obtain the optimized control strategy. The target operating parameters include hydrogen flow rate, air flow rate, stack cooling temperature, and output voltage. The control strategy is converted into a standard control signal, and the target operating parameters are adjusted according to the standard control signal.

9. A training system for a power supply life prediction model of a high-voltage grounding wire repair robot, characterized in that, The training system includes: The first acquisition module is used to acquire historical multidimensional behavior data of the high-voltage grounding wire repair robot, wherein the historical multidimensional behavior data includes overall energy data, robot behavior data, fuel cell operation data and environmental status data; The feature filtering module is used to combine Pearson correlation coefficient analysis and mutual information entropy method to perform dual-criteria feature filtering on the historical multidimensional behavioral data to obtain the core features related to fuel cell degradation. The entropy weight quantization module is used to perform entropy weight quantization based on the information entropy of the core features to obtain the core weighted features. The pre-training module is used to train the general prediction model to be trained based on the core weighted features to obtain a trained general prediction model. The general prediction model is obtained by introducing a multi-head attention mechanism on the basis of the time series prediction model. The energy analysis module is used to perform energy analysis based on the historical multidimensional behavior data to obtain energy patterns corresponding to different work behaviors, wherein the energy patterns include energy demand patterns and energy consumption patterns. The transfer training module is used to adapt and train the general prediction model based on the historical multidimensional behavior data and the energy pattern using a transfer learning strategy, so as to obtain the power life prediction model corresponding to each type of operation behavior.

10. A power supply life prediction system for a high-voltage grounding wire repair robot, characterized in that, The prediction system includes: The second acquisition module is used to acquire the core behavioral data at the current moment, wherein the core behavioral data is the core features related to fuel cell degradation; The behavior matching module is used to perform behavior matching based on the core behavior data to obtain the current behavior pattern of the high-voltage grounding wire repair robot's current operation behavior; The lifespan prediction module is used to call a pre-trained power supply lifespan prediction model based on the current behavior pattern, and input the core behavior data into the power supply lifespan prediction model to predict the power supply lifespan and obtain the current prediction result.

Images