Control method for inspection robot in narrow space of power station

Abstract

The invention relates to the technical field of robot control, and discloses a power station narrow space inspection robot control method. According to the method, when a preset inspection triggering condition is met, a historical operation state of a target inspection robot at a specified historical moment is obtained, the historical operation state comprises historical accumulated operation duration and historical environment parameters, and meanwhile, a real-time operation state is obtained, and the real-time operation state comprises real-time accumulated operation duration and real-time environment parameters. Through analysis of historical and real-time operation states, an operation state deviation vector is constructed, and a corresponding operation deviation measurement set is retrieved. When the set distribution dispersion is larger than or equal to a dispersion threshold value, the ratio of the historical duration from the storage starting moment to the specified historical moment to the current duration from the storage starting moment to the current moment is calculated and serves as a control parameter adjusting proportion, and then control parameters of the target inspection robot are updated according to the proportion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot control, in particular to a control method of a narrow-space inspection robot in a power station. BACKGROUND

[0002] In modern power stations, narrow spaces are widely present, such as cable tunnels, equipment interlayers, and the interiors of control cabinets. These narrow spaces are full of various types of equipment, and various types of cables and pipelines are crisscrossed therein, with valves, instruments, and other equipment densely distributed therein, forming a very complex environment. For example, in the cable tunnels of some large-scale substations, the space is narrow and winding, the height is usually only enough for one person to pass through by bending over, and the interior is full of high-voltage cables, communication cables, and the like, with a high humidity and possibly with accumulated water, and a slight carelessness can cause a safety accident.

[0003] Traditional manual inspection faces many difficulties in such narrow spaces. First, the limited space makes it extremely inconvenient for the inspection personnel to move around, and it is difficult for them to freely check the equipment in all directions. For example, in the equipment interlayer, the inspection personnel may need to turn sideways, bend over, or even crawl to reach some equipment positions, which is extremely difficult to operate. Second, manual inspection is low in efficiency, and in the face of numerous narrow spaces and equipment in the power station, a large amount of time is consumed for manual inspection one by one, which cannot meet the needs of efficient operation of the power station. According to statistics, it may take several days to complete a comprehensive manual inspection in a medium-sized power station, and during this period, the operating state of the power station equipment cannot be monitored in a timely manner. Third, there are many safety risks in narrow spaces, such as the accumulation of harmful gases, oxygen deficiency, high temperature, and other harsh environments, which threaten the safety of the inspection personnel. In the interior of the control cabinet of some old power stations, there may be harmful gases due to poor ventilation, and if the inspection personnel do not take protective measures when entering, poisoning incidents are likely to occur.

[0004] In order to overcome the shortcomings of manual inspection, traditional inspection robots have been introduced into the inspection work of power stations. However, the traditional inspection robots also have many drawbacks in narrow spaces. First, they have poor flexibility. Most traditional inspection robots are large in size and not compact in structure, and it is difficult for them to move flexibly in narrow passages and complex equipment layouts, and they often cannot reach the designated positions for inspection. For example, some wheeled inspection robots have a large turning radius when encountering bends or obstacles in the cable tunnel, and cannot pass through smoothly, seriously affecting the normal development of the inspection work.

[0005] Second, they have weak adaptability to environmental changes. The environmental parameters of the narrow spaces in power stations, such as temperature, humidity, and electromagnetic interference, change frequently, and the sensors and control systems of traditional inspection robots are difficult to quickly and accurately adapt to these changes. When the temperature suddenly rises or the electromagnetic interference increases, the sensors of the robot may have data deviation, leading to misjudgment of the equipment state and failure to timely discover equipment fault hazards.

[0006] Control parameters are difficult to adjust dynamically. Traditional inspection robots typically have pre-set control parameters, making it difficult to dynamically adjust them based on real-time operating conditions and environmental changes. At different times, the operating load of power plant equipment varies, and environmental parameters change accordingly. Traditional robots cannot optimize their operating strategies based on these changes, resulting in poor inspection performance and failing to fully realize their intended purpose. Summary of the Invention

[0007] The purpose of this invention is to provide a control method for a robot inspecting confined spaces in a power plant, in order to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides a control method for a power plant confined space inspection robot, the method comprising: When the preset inspection triggering conditions are met, the historical running status of the target inspection robot at a specified historical moment is obtained, where the historical running status includes the historical cumulative running time and historical environmental parameters. The real-time operating status of the target inspection robot is obtained, including the real-time cumulative running time and real-time environmental parameters. Based on the historical operating status and the real-time operating status, an operating status deviation vector is constructed. The operating deviation metric set corresponding to the operating status deviation vector is retrieved. When the distribution dispersion of the operating deviation metric set is greater than or equal to the dispersion threshold, the ratio of the historical duration from the storage start time to the specified historical time to the current duration from the storage start time to the current time is calculated and used as the control parameter adjustment ratio. The control parameters of the target inspection robot are updated according to the control parameter adjustment ratio.

[0009] Preferably, the step of obtaining the historical operating status of the target inspection robot at a specified historical moment includes: obtaining multiple environmental parameter sequences of the target inspection robot; for each environmental parameter sequence, calculating the dynamic time warping distance between the environmental parameter sequence and the historical benchmark sequence; determining the alienated runtime corresponding to each environmental parameter sequence based on the dynamic time warping distance; extracting the minimum value among all alienated runtimes, multiplying the minimum value by a preset coefficient, and obtaining the preset inspection triggering condition.

[0010] Preferably, the step of obtaining the historical operating status of the target inspection robot at a specified historical moment further includes: real-time monitoring of the amount of operating data of the target inspection robot; calculating the instantaneous growth rate of the amount of operating data; and automatically triggering the acquisition of the historical operating status of the target inspection robot at a specified historical moment when the instantaneous growth rate exceeds a preset rate threshold.

[0011] Preferably, the step of calculating the dynamic time-warped distance between each environmental parameter sequence and the historical reference sequence includes: when the environmental parameter sequence is the original sensor reading, retrieving multiple historical sensor reading sequences and corresponding runtime sequences from the historical database; performing dynamic time warping calculation on each historical sensor reading sequence and the current sensor reading sequence to obtain the warped path distance for each sequence pair; collecting all warped path distances to form a distance set; filtering out a subset of warped path distances greater than a preset distance threshold from the distance set; and performing mode calculation on the runtime sequences corresponding to the warped path distance subsets to obtain the heterogeneous runtime.

[0012] Preferably, the step of calculating the dynamic time warping distance between each environmental parameter sequence and the historical benchmark sequence includes: when the environmental parameter sequence is a fusion feature, calling the feature extraction model associated with the fusion feature, wherein the feature extraction model is constructed based on a pre-trained neural network; processing historical sensor data through the feature extraction model to obtain a historical fusion feature sequence; calculating the cosine similarity between the historical fusion feature sequence and the current fusion feature sequence; and deriving the alienation runtime based on the deviation between the cosine similarity and a preset similarity threshold.

[0013] Preferably, the method further includes: configuring the network structure of the feature extraction model and specifying the parameters of the convolutional layer and the fully connected layer; collecting multiple sets of training data, each set of training data including sensor reading sequences and corresponding fused feature labels; optimizing the weights of the feature extraction model using the gradient descent algorithm; and binding the optimized feature extraction model with specific fused features and environmental parameters.

[0014] Preferably, the step of retrieving the set of operation deviation metrics corresponding to the operation state deviation vector includes: decomposing the operation state deviation vector into a time deviation component and an environmental deviation component; encrypting the time deviation component and the environmental deviation component using a secure hash algorithm to generate a digital digest; sending the digital digest to the blockchain network and retrieving matching operation deviation records through a deployed smart contract; obtaining an initial set of operation deviation metrics from the results returned by the smart contract; and performing a moving average filtering process on the initial set of operation deviation metrics to remove outliers, thereby obtaining the final set of operation deviation metrics.

[0015] Preferably, the step of sending the digital digest to the blockchain network includes: deploying a retrieval smart contract in the blockchain network, the contract containing matching logic based on deviation components; calling the smart contract through a distributed node client, passing in the digital digest as an input parameter; the smart contract querying records in the distributed ledger that meet the deviation conditions and returning encrypted query results; and decrypting the query results using a private key to obtain an initial running deviation metric set.

[0016] Preferably, the step of updating the control parameters of the target inspection robot according to the control parameter adjustment ratio includes: inputting the control parameter adjustment ratio into the fuzzy logic controller; the fuzzy logic controller calculating the control output according to the ratio value, including speed adjustment amount and direction adjustment amount; and sending the speed adjustment amount and direction adjustment amount to the motion control unit of the inspection robot to adjust the actual motion trajectory.

[0017] Preferably, the method further includes: after constructing the operating state deviation vector, performing minimum-maximum scaling on each component of the operating state deviation vector to adjust the numerical range of all components to between 0 and 1; and using the scaled operating state deviation vector for subsequent retrieval steps.

[0018] Compared with the prior art, the beneficial effects of the present invention are: To accurately grasp the operational status of inspection robots, the historical operational status of the target inspection robot at a specified historical moment is obtained, including historical cumulative runtime and historical environmental parameters. Simultaneously, its real-time operational status is acquired, encompassing real-time cumulative runtime and real-time environmental parameters. These two data points are compared to construct an operational status deviation vector. This process allows maintenance personnel to gain a deep understanding of the robot's operational status and gain insights into its working condition from multiple dimensions. For example, by comparing and analyzing temperature and humidity data from historical and real-time environmental parameters, the impact of environmental changes on robot operation can be identified in a timely manner, enabling proactive measures to ensure stable robot operation and effectively avoid inspection errors caused by abnormal robot conditions.

[0019] This method demonstrates significant advantages in handling complex environmental changes. The confined environment of a power plant is complex and constantly changing, with environmental parameters such as temperature, humidity, and electromagnetic interference constantly shifting. Based on the acquired environmental parameters, this method can dynamically adjust control parameters. When a sudden increase in ambient humidity is detected, the robot's movement speed and sensor operating modes can be adjusted accordingly to prevent rusting of robot components or distortion of sensor data caused by humid environments. In this way, the robot can operate stably in various complex environments, improving its adaptability to different working conditions and ensuring the continuity and accuracy of inspection work.

[0020] This control method demonstrates significant effectiveness in reducing operation and maintenance costs. Traditional inspection robot maintenance often lacks specificity, easily leading to resource waste. This method, however, rationally adjusts control parameters based on operational deviations, avoiding over-maintenance. When the dispersion of the operational deviation metric set is greater than or equal to a dispersion threshold, the ratio of the historical duration from the storage start time to a specified historical time to the current duration from the storage start time to the current time is calculated and used as the control parameter adjustment ratio to update the control parameters of the target inspection robot. This allows for precise resource allocation and maintenance based on the robot's actual operating conditions, reducing unnecessary maintenance work and resource consumption, lowering the input of human, material, and financial resources, improving operation and maintenance efficiency, and achieving effective control of operation and maintenance costs.

[0021] This method plays a crucial role in improving the reliability and safety of inspections. Through precise operational status analysis and reasonable adjustment of control parameters, it effectively reduces the occurrence of misjudgments and omissions. During inspections, the robot can more accurately detect potential equipment faults and issue timely warnings, saving valuable time for the maintenance and repair of power plant equipment. This not only ensures the stable operation of power plant equipment but also reduces the risk of safety accidents caused by equipment failures, protecting the lives of power plant personnel and building a solid defense for the stable operation of the power plant. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the working principle of the power plant confined space inspection robot control method described in this invention. Figure 2 A sub-flowchart for obtaining historical running status; Figure 3 The sub-flowchart for calculating the dynamic time warped distance. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1This invention provides a control method for a power plant inspection robot in confined spaces. The method includes: achieving precise control of the inspection robot by comprehensively analyzing historical and real-time operating states. When a preset inspection trigger condition is met, the system acquires the historical operating state of the target inspection robot at a specified historical moment, including historical cumulative runtime and historical environmental parameters; simultaneously, it acquires the real-time operating state of the target inspection robot, including real-time cumulative runtime and real-time environmental parameters. Based on the historical and real-time operating states, an operating state deviation vector is constructed, which quantifies the differences in operating states. Subsequently, the system retrieves the operating deviation metric set corresponding to the operating state deviation vector. When the dispersion of the operating deviation metric set is greater than or equal to a dispersion threshold, the system calculates the ratio of the historical duration from the storage start time to the specified historical moment to the current duration from the storage start time to the current moment. This ratio is used as the control parameter adjustment ratio. Finally, the control parameters of the target inspection robot are updated according to the control parameter adjustment ratio, thereby optimizing the robot's operating performance.

[0025] Example 1: See Figure 2 In acquiring the historical operating status of the target inspection robot at a specified historical moment, the system obtains multiple environmental parameter sequences of the target inspection robot. These environmental parameter sequences are continuously generated by various sensors deployed in the confined space of the power plant, including temperature sensors, humidity sensors, gas concentration sensors, and vibration sensors. Each environmental parameter sequence represents a set of readings of a specific environmental parameter within a continuous time period. For each environmental parameter sequence, the system initiates a dynamic time warping algorithm to calculate the dynamic time warping distance between the environmental parameter sequence and a pre-stored historical benchmark sequence. The historical benchmark sequence is a typical environmental data pattern selected from long-term stable operation phases. The calculation process of the dynamic time warping distance involves constructing a cumulative cost matrix and finding the optimal curvature path between the two sequences through dynamic programming, thereby eliminating the influence of stretching and deformation on the time axis on the distance measurement. Based on the calculated dynamic time warp distance, the system executes the derivation logic for the alienated runtime. The value of the dynamic time warp distance is positively correlated with the degree of abnormal deviation of environmental parameters. The system maintains a distance-duration mapping table, which is established by analyzing historical abnormal data. This mapping table maps the dynamic time warp distance value to an estimated abnormal duration, i.e., the alienated runtime. After processing all environmental parameter sequences, the system extracts the minimum value among all alienated runtimes. This minimum value represents the most significant short-term abnormal indicator in the current environment. This minimum value is multiplied by a preset coefficient, which is jointly determined by domain experts based on power plant operation safety criteria and robot performance indicators. The calculation result is the threshold for the preset inspection trigger condition.

[0026] The system synchronously performs real-time monitoring of the target inspection robot's operational data. This data includes raw sensor readings, processed feature data, control command logs, and system status information. Data volume is statistically measured in packets or bytes within a fixed time window. The system calculates the instantaneous growth rate of the operational data by comparing the difference in data volume between adjacent time windows and dividing by the time interval. This calculation method captures abrupt changes in the data stream. When the instantaneous growth rate exceeds a preset threshold, which is dynamically adjusted based on the statistical distribution of historical operational data, the system immediately triggers an interrupt signal and automatically initiates the process of acquiring the target inspection robot's historical operational status at a specified historical moment. The acquisition of environmental parameter sequences is integrated into the robot's data acquisition module. Multi-source sensor data is transmitted to the central processing unit via a unified communication protocol, and the data sequences are aligned and cached according to timestamps. The construction of historical baseline sequences employs offline machine learning methods to cluster representative sequence patterns from a large amount of normal operational data. The implementation of the dynamic time warping algorithm employs an optimized computational strategy to reduce the computational overhead of real-time processing. The derivation of the heterogeneous runtime introduces a confidence verification mechanism to avoid misjudgments caused by instantaneous sensor noise. The setting of preset coefficients takes into account the robot's mechanical wear cycle and the periodic changes in environmental factors. The generation of preset inspection trigger conditions is a dynamic and continuous process.

[0027] Monitoring of operational data volume relies on high-precision timers and data counters in the embedded system. The instantaneous growth rate is calculated using a sliding window algorithm to smooth random fluctuations. The dynamic adjustment strategy for the preset rate threshold is based on a feedback mechanism in control theory; when the system is under high load, the threshold is automatically increased to prevent oversensitivity. The mechanism for automatically triggering the acquisition of historical operating states is designed as a low-latency interrupt service routine to ensure timely status acquisition. In the specific implementation of calculating the dynamic time warping distance between the environmental parameter sequence and the historical baseline sequence, the system accesses a historical database, which uses a time-series database architecture to optimize time range queries. Multiple historical sensor reading sequences and their corresponding runtime sequences are retrieved; the historical sensor reading sequences are stored by device number and time partition. Dynamic time warping calculation is performed on each historical sensor reading sequence and the current sensor reading sequence. The calculation process uses a fast algorithm for constrained curved paths to limit computational complexity. The warped path distance for each sequence pair is obtained, and the warped path distance is normalized to eliminate the influence of sequence length. All warped path distances are collected to form a distance set, which is stored in a temporary in-memory database for rapid analysis. A subset of regularized path distances greater than a preset distance threshold is selected from the distance set. The preset distance threshold is dynamically calculated based on the quantiles of the distance distribution using an adaptive algorithm. The mode of the runtime sequence corresponding to the regularized path distance subset is calculated using a hash table frequency counting method combined with secondary verification. The alienated runtime is then obtained and output to the decision logic unit.

[0028] The historical database index structure is optimized for time-series queries, and the retrieval of historical sensor reading sequences supports fuzzy matching to address sensor clock drift. The implementation of dynamic time warping calculation includes boundary condition handling mechanisms, and the collection process for warped path distances performs data integrity verification. Distance set filtering uses streaming processing to avoid memory overflow, and the adaptive algorithm for preset distance thresholds references the moving average of recent distance values. The mode calculation result uses a time rationality filter to remove obvious outliers, and the output format of the variable runtime is compatible with downstream module interfaces. The module for real-time monitoring of the target inspection robot's operational data volume runs on a real-time operating system, and the statistics of operational data volume cover application layer and network layer data packets. The instantaneous growth rate of operational data volume is calculated, and a Kalman filter is introduced to preprocess the raw data to reduce noise interference. When the instantaneous growth rate exceeds a preset rate threshold, the system generates an event log to record the triggered event for post-event analysis. The system automatically triggers the acquisition of the target inspection robot's historical operating status at a specified historical moment, and resource availability checks are performed before the status acquisition process begins. The accuracy of data volume statistics is configurable to adapt to different network environments, and the alarm conditions for the instantaneous growth rate support combinational logic expressions. The preset rate threshold adjustment strategy allows for manual intervention, while the automatic triggering mechanism ensures system stability under abnormal data impacts. Persistent storage of environmental parameter sequences utilizes compression algorithms to save storage space, and the historical baseline sequence update strategy supports incremental learning. A configurable balance is achieved between the calculation accuracy and computational resource consumption of the dynamic time-warped distance, and the derivation model for heterogeneous runtime supports online updates. The preset coefficient adjustment interface provides visual tools for easy operation by maintenance personnel, and the evaluation cycle of preset inspection trigger conditions is synchronized with the robot's control cycle. Data collection points are distributed across various robot subsystems, and the instantaneous growth rate monitoring panel is integrated into a centralized management platform. The initial value of the preset rate threshold is derived from the system's factory test data, and the status acquisition process after automatic triggering features a recoverable design.

[0029] Example 2: See Figure 3In the specific implementation of calculating the dynamic time warping distance between each environmental parameter sequence and the historical baseline sequence, when the environmental parameter sequence is the raw sensor reading, the system initiates a multi-stage data processing flow. The raw sensor readings come directly from the physical sensor modules installed in the confined space of the power station, and these readings constitute a time-series data stream. The system retrieves multiple historical sensor reading sequences and corresponding runtime sequences from the historical database. The historical database is built on a distributed file system and uses a columnar storage format to optimize the reading performance of time-series data. The historical sensor reading sequences are indexed using a composite index based on device identifier, sensor type, and timestamp. The corresponding runtime sequence records the continuous running time of the robot corresponding to each historical data segment. The runtime sequence and the sensor reading sequence are strictly associated through a globally unique transaction identifier. Dynamic time warping calculation is performed on each historical sensor reading sequence and the current sensor reading sequence. The dynamic time warping calculation is executed by a dedicated digital signal processing coprocessor. The calculation process applies an improved dynamic time warping algorithm with constrained bending windows, which balances calculation accuracy and computational efficiency by limiting the maximum curvature of the time axis. The computation process generates a cumulative cost matrix. A backtracking method is used to find a regularization path that minimizes the cumulative cost. This regularization path defines the optimal nonlinear alignment between two sequences. The regularization path distance for each sequence pair is obtained. This regularization path distance is a scalar value representing the overall difference between the two sequences after optimal regularization. The regularization path distance is normalized to eliminate measurement bias caused by inconsistent sequence lengths.

[0030] All regularized path distances are collected to form a distance set, which is stored in a cache database configured with an automatic expiration policy to manage the data's lifecycle. A subset of regularized path distances greater than a preset threshold is selected from this set. This preset threshold is dynamically determined by an adaptive thresholding algorithm, which analyzes the statistical distribution characteristics of the distance set and calculates a critical value based on information from a historical anomaly pattern database. The selection operation uses a predicate-based filtering technique, retaining only regularized path distances that significantly deviate from the normal pattern. The mode is calculated for the runtime sequences corresponding to the regularized path distance subsets. This mode calculation employs an improved majority voting algorithm, capable of handling multi-peak distributions and identifying the most representative runtime values. A weighting mechanism is introduced in the mode calculation process; the weight value is positively correlated with the magnitude of the regularized path distance, with larger regularized path distances having higher voting weights in the mode calculation. The alienated runtime is then obtained and output as a key time metric to the decision subsystem. The historical database retrieval interface supports both batch and streaming query modes. Batch queries are used to load large amounts of historical data during the initialization phase, while streaming queries are used to process continuously generated data segments in real time. Retrieval conditions for historical sensor reading sequences include time range filtering, sensor type filtering, and data quality filtering. Data quality filtering excludes sequences containing a large number of missing values ​​or obvious errors based on predefined data integrity rules. The retrieval of corresponding runtime sequences ensures data consistency through distributed transactions, and the time resolution of the runtime sequence is synchronized with the sensor reading sequence. The implementation of dynamic time warping calculation is optimized for embedded environments, and the algorithm code uses memory-efficient data structures to reduce the storage overhead of intermediate variables. The calculated results of the warped path distance are rounded and pruned to ensure that the output value is within a reasonable numerical range.

[0031] The distance set management employs a sliding time window mechanism. New regularized path distances are continuously added to the distance set, while outdated regularized path distances are automatically removed, ensuring that the distance set always reflects the data characteristics of the most recent period. An adaptive algorithm with a preset distance threshold is re-executed at fixed time intervals, and the update process of the preset distance threshold has a smooth transition characteristic to avoid numerical jumps. The generation of a subset of regularized path distances triggers a conditional event, which notifies subsequent processing modules to start the mode calculation task. The mode calculation of the runtime sequence is executed in a separate calculation thread, and the data of the runtime sequence is pre-sorted to improve computational efficiency. The output format of the heterogeneous runtime includes a timestamp and confidence information; the confidence information reflects the reliability of the mode calculation result. The system integrates a data quality monitoring step when processing raw sensor readings. Raw sensor readings undergo a preprocessing pipeline before entering the dynamic time regularization calculation, and the preprocessing pipeline includes signal filtering, outlier removal, and data normalization steps. Signal filtering employs digital filters to eliminate high-frequency noise. Outlier removal uses statistical rules to identify and remove data points that significantly deviate from the normal range. Data normalization maps readings from different sensors to a unified numerical scale. The historical database architecture supports horizontal scaling, and its sharding strategy divides data according to the time dimension, storing newer data on high-speed storage devices and archiving older data on low-cost storage media. The compression algorithm for historical sensor reading sequences is specifically optimized for sensor data patterns, achieving a high compression ratio while preserving data characteristics.

[0032] The parallel implementation of dynamic time warping computation leverages the computing power of multi-core processors. The task scheduler for dynamic time warping computation distributes computational tasks for multiple sequence pairs across different processing cores for simultaneous execution. The collector for warping path distance aggregates results from various computing nodes, and the transmission protocol for warping path distance ensures data integrity and order. The storage structure of the distance set uses a self-balancing binary search tree, and the index of the distance set is built on the numerical values ​​of the warping path distances to achieve fast range queries. The computation engine for preset distance thresholds integrates a machine learning model, and the model training uses historical normal and abnormal data as training samples. The iterator for warping path distance subsets implements a lazy loading mode, preventing large result sets of warping path distance subsets from being loaded into memory all at once. The mode calculator for runtime sequences handles boundary cases, returning a special error code when the runtime sequence is empty. The validation module for heterogeneous runtimes checks the reasonableness of the output values, and the value range of heterogeneous runtimes is limited to the maximum and minimum time range allowed by the system.

[0033] The acquisition frequency of raw sensor readings is dynamically adjusted according to the robot's operating status. The acquisition frequency is increased when the robot is moving at high speed and decreased when the robot is stationary to save energy. The historical baseline sequence update strategy uses incremental learning, periodically incorporating new normal data patterns to maintain the baseline's current status. The path constraint parameters of the dynamic time warping algorithm are experimentally determined; their values ​​affect the algorithm's tolerance to sequence time distortion. The normalization method for the warped path distance is min-max normalization, with normalization parameters learned from training data. A visualization tool for the distance set helps operations personnel understand data distribution, and statistical reports on the distance set are generated periodically for system performance evaluation. A preset distance threshold alarm mechanism notifies the administrator when the threshold fluctuates abnormally, and the modification log for the preset distance threshold records all change history. The serialization format of the warped path distance subset adopts a cross-platform data exchange standard, and persistent storage of the warped path distance subset supports fault recovery. A runtime long-sequence time alignment algorithm handles sensor clock asynchrony issues, and a runtime long-sequence interpolation method fills in small-scale data gaps. The hash table size for mode calculation is dynamically adjusted based on the data volume, and concurrency control for mode calculation prevents multi-threaded access conflicts. A caching mechanism for alienated runtime stores frequently used calculation results, and the lifecycle of alienated runtime is bound to a sequence of related environment parameters.

[0034] Example 3: In the implementation of calculating the dynamic time warping distance between the environmental parameter sequence and the historical baseline sequence, when the environmental parameter sequence is a fused feature, the system calls the feature extraction model associated with the fused feature. The feature extraction model is constructed based on a pre-trained neural network. The historical sensor data is processed by the feature extraction model to obtain the historical fused feature sequence, which is a high-dimensional representation of the original sensor data after nonlinear transformation. The cosine similarity between the historical fused feature sequence and the current fused feature sequence is calculated. The cosine similarity measures the degree of similarity between the two feature vectors in direction. Based on the deviation of the cosine similarity from a preset similarity threshold, the alienation runtime is derived. The alienation runtime reflects the length of time the system behavior deviates from the normal pattern.

[0035] The cosine similarity is calculated using the following mathematical expression: in: This represents the cosine similarity value between the historical fusion feature sequence and the current fusion feature sequence, and this value ranges from -1 to 1. This represents the feature value of the i-th dimension in the historical fusion feature sequence, which is generated by the feature extraction model through processing historical sensor data. This represents the feature value of the i-th dimension in the current fused feature sequence, which is generated in real time during the operation of the inspection robot. (Symbol) Define the total number of dimensions of the fused feature vectors; this parameter is determined during the feature extraction model design phase.

[0036] The system configures the network structure of the feature extraction model, specifying the parameters for convolutional and fully connected layers. Convolutional layer parameters include the number of filters, filter size, stride, and padding. Fully connected layer parameters include the number of neurons, activation function type, and dropout rate. Multiple sets of training data are collected, each including sensor reading sequences and corresponding fused feature labels, covering various operating conditions and environments. Gradient descent is used to optimize the weights of the feature extraction model, iteratively adjusting network parameters to minimize the loss function. The optimized feature extraction model is then bound to specific fused features and environmental parameters, establishing a mapping relationship between the model and data types. The network structure of the feature extraction model is implemented using a deep learning framework. Convolutional layers extract local spatiotemporal features from sensor data, while fully connected layers perform nonlinear combinations of features. The number of filters determines the number of channels in the feature map, filter size affects the receptive field size, stride controls the feature map downsampling rate, and padding maintains the spatial size of the feature map. The number of neurons in the fully connected layer matches the dimension of the fused features, the activation function introduces nonlinear transformation capabilities, and the dropout rate prevents overfitting. Training data acquisition was accomplished through automated scripts. Sensor reading sequences were preprocessed and paired with manually labeled fused feature tags. The gradient descent algorithm was executed using backpropagation to calculate the gradient, and the learning rate was adaptively adjusted to balance convergence speed and stability.

[0037] The feature extraction model's API is encapsulated as a standardized service, and historical sensor data loading employs a streaming processing model. Batch generation of historical fused feature sequences improves efficiency, while real-time calculation of the current fused feature sequence ensures timeliness. Cosine similarity calculation is optimized to matrix operations, leveraging hardware acceleration to enhance computation speed. The preset similarity threshold is set based on statistical quantiles, with a threshold update mechanism adapting to changes in data distribution. The derivation of the heterogeneous runtime establishes a similarity-time mapping model, trained through regression analysis. The feature extraction model's deployment environment is configured with a graphics processor for accelerated computation, and depthwise separable convolutions are used to reduce the number of parameters in the convolutional layers. The initialization of fully connected layers uses the Xavier method to maintain gradient stability, and augmentation techniques for training data increase sample diversity. Normalization of sensor reading sequences eliminates the influence of dimensions, and fused feature labels are generated using an autoencoder. The Adam optimizer is chosen as a variant of the gradient descent algorithm, and a weight decay term is incorporated into the loss function to prevent overfitting. Metadata bound to the model records version information, and the selection of specific fused features is based on feature importance ranking.

[0038] The system implements online updates for the feature extraction model, and a visualization tool for convolutional layer activation assists in model debugging. Batch normalization of fully connected layers accelerates the training process, and a training data quality monitoring module filters out abnormal samples. The gradient descent algorithm's convergence criterion includes an early stopping mechanism, and the optimized feature extraction model is exported in a standard format. The configuration interface for binding relationships supports graphical operation, and the classification structure of environmental parameters is stored in a structured manner. Historical sensor data is stored in compressed binary format, and the index of historical fused feature sequences constructs a feature space index. The caching mechanism for current fused feature sequences reduces redundant computation, and parallel computation of cosine similarity utilizes multi-threading technology. Dynamic adjustment of preset similarity thresholds is based on sliding window statistics, and the output interface for heterogeneous runtime is integrated with the decision system. The feature extraction model's monitoring panel displays its running status, and the visualization of convolutional layer feature maps helps understand model behavior. A version management system for training data records data changes, and a labeling tool for sensor reading sequences supports semi-automatic labeling. The distributed implementation of the gradient descent algorithm handles large-scale data, and the evaluation metrics for the feature extraction model include accuracy and recall. Validation tests of binding relationships ensure model-data compatibility, and the network structure configuration file uses a human-readable format. The number of filters is determined experimentally to select the optimal value, and the activation function is chosen based on the feature distribution characteristics. Training data partitioning maintains temporal consistency, and the quality of fused feature labels impacts model performance. The learning rate scheduling for the gradient descent algorithm employs a cosine annealing strategy, and the gradient change trend is monitored during weight optimization. Dependency management for model binding resolves version conflicts, and the extraction process for specific fused features is standardized. The inference engine of the feature extraction model optimizes the computational graph structure, and hardware acceleration of convolutional layers utilizes a dedicated instruction set. Sparsity processing of fully connected layers reduces computational overhead, and the training data management platform supports collaborative annotation. Gradient pruning in the gradient descent algorithm prevents the explosion problem, and encryption of the feature extraction model protects intellectual property rights. Automated testing of binding relationships covers boundary conditions, and metadata management of environmental parameters ensures a robust traceability chain. Strict rules are enforced for verifying the integrity of historical sensor data, and similarity searches of historical fused feature sequences respond quickly. The current fused feature sequence generation pipeline is efficient, and the numerical stability of cosine similarity handles boundary values. The adaptive algorithm with preset similarity thresholds is robust, and the computational accuracy for varying runtime meets application requirements.

[0039] Example 4: In the process of retrieving the set of operating deviation metrics corresponding to the operating state deviation vector, the system decomposes the operating state deviation vector into a time deviation component and an environmental deviation component. The time deviation component represents the cumulative difference in robot runtime, while the environmental deviation component reflects the instantaneous changes in the environmental parameters of the robot. A secure hash algorithm is used to encrypt the time deviation component and the environmental deviation component. The secure hash algorithm uses the SHA-256 standard to generate a fixed-length digital digest. The digital digest is sent to a blockchain network, which is a decentralized ledger system composed of multiple distributed nodes. Matching operating deviation records are retrieved through a deployed smart contract, which contains pre-programmed query logic. An initial set of operating deviation metrics is obtained from the results returned by the smart contract, containing the original retrieved data. A moving average filter is applied to the initial set of operating deviation metrics, using a sliding window of size 5 to calculate a continuous average. Outliers are removed to obtain the set of operating deviation metrics, which is used for subsequent distribution dispersion calculations. The extraction of the time deviation component is achieved through time series analysis, while the calculation of the environmental deviation component is based on multivariate statistical analysis. The input data format for the secure hash algorithm is standardized and preprocessed, and timestamps are added during the digital digest generation process to prevent replay attacks. The choice of blockchain network considers transaction processing speed and consensus mechanisms, and smart contract deployment uses an Ethereum Virtual Machine-compatible environment. The initial data structure design of the deviation metric set supports fast indexing, and the parameters of the moving average filter are adjusted according to the data sampling frequency. Outlier removal employs box plot statistical methods, and the storage format of the deviation metric set is optimized for space efficiency.

[0040] The specific steps for sending the digital digest to the blockchain network include deploying a retrieval smart contract within the blockchain network. This smart contract is written in Solidity and compiled into bytecode. The smart contract contains matching logic based on deviation components, implementing a multi-condition query algorithm. The smart contract is invoked through a distributed node client, which is configured with node discovery and connection management functions. The digital digest is passed as input parameters, and the encoding of the input parameters conforms to the ABI specification. The smart contract queries the distributed ledger for records that meet the deviation conditions, using a state tree index to accelerate data access. It returns encrypted query results, protected by an asymmetric encryption algorithm. The query results are decrypted using a private key stored in a hardware security module. An initial set of running deviation metrics is obtained, and after format conversion, it is fed into the processing pipeline. Referring to Table 1, the code for retrieving smart contracts undergoes formal verification to ensure logical correctness, and the matching logic supports fuzzy queries to adapt to data uncertainty. The distributed node client implementation includes a retry mechanism to handle network fluctuations, and the transmission of digital digests uses a secure channel to prevent eavesdropping. Query optimization for smart contracts reduces gas consumption, and the distributed ledger's index structure supports range queries. The encryption key for returned results is dynamically generated, and the security protocol for private key management conforms to financial-grade standards. Deserialization of the initial runtime deviation metric set checks data integrity.

[0041] Table 1: Parameter Table for Querying Operational Deviation Records The blockchain network's node configuration employs a Byzantine fault-tolerant consensus algorithm, and the storage layout for retrieving smart contracts optimizes gas consumption. The matching logic supports queries with composite conditions, and the weighting coefficients for time and environmental deviation components are configurable. A load balancing algorithm for distributed node clients distributes query pressure, and random numbers are added to the generation of digital digests to increase entropy. The smart contract query interface design supports batch operations, and data partitioning of the distributed ledger improves parallel processing capabilities. Return result compression reduces network transmission volume, and performance optimization for private key decryption operations utilizes a dedicated cryptographic processor. The initial running deviation metric set's verification rules check data boundaries. The secure hash algorithm implementation uses a certified cryptographic library, and normalization processing for time and environmental deviation components eliminates the influence of dimensions. The digital digest transmission protocol ensures end-to-end integrity, and the blockchain network's transaction management handles concurrent conflicts. The smart contract upgrade mechanism supports seamless replacement, and the distributed ledger's snapshot function accelerates query response. The filtering conditions for the initial running deviation metric set are dynamically configurable, and the window size for moving average filtering can be adaptively adjusted. The sensitivity parameters for outlier detection support online updates, and the serialization format of the running deviation metric set is backward compatible.

[0042] The blockchain network's monitoring system tracks node status and retrieves detailed operation history from smart contract logs. Matching logic test cases cover boundary conditions, and the connection pool of distributed node clients manages network resources. A digital digest caching mechanism reduces redundant computation, and the smart contract query optimizer selects the optimal execution plan. The distributed ledger's storage engine supports horizontal scaling, and pagination of returned results supports large datasets. Secure storage of private keys uses multi-factor authentication, and sampling statistics of the initial runtime deviation metric set generate metadata. The calculation accuracy of the time deviation component reaches millisecond levels, and feature engineering of the environmental deviation component extracts key indicators. Salt addition in the secure hash algorithm enhances collision resistance, and the digital digest verification mechanism prevents tampering. The blockchain network's smart contract event system supports asynchronous notifications, and the matching logic's scoring algorithm sorts results. Distributed node client failover ensures high availability, and smart contract state rollback handles exceptional situations. The distributed ledger's Merkle tree proof verifies data integrity, and data anonymization of returned results protects privacy. A private key rotation strategy periodically updates key materials, and an archiving strategy for the initial runtime deviation metric set manages the data lifecycle. The decomposition algorithm for the runtime deviation vector has linear time complexity, and the weighting coefficients of the time deviation component and the environmental deviation component are obtained through machine learning training. Parallel computation of the secure hash algorithm utilizes multi-core processors, and the digital digest is stored using a distributed hash table. The cross-chain interoperability protocol of the blockchain network supports multi-chain queries, and the oracle of the smart contract integrates external data sources. The protocol stack of the distributed node client supports multiple network protocols, and the transformation of the initial runtime deviation metric set preserves data semantics. Boundary processing for the moving average filter uses mirror filling, and the statistical model for outlier removal is based on the Gaussian distribution assumption. A visualization tool for the runtime deviation metric set assists in result analysis.

[0043] The consensus parameters of the blockchain network are adjusted according to the network size, and access control lists manage permissions for smart contracts. The index structure for matching logic uses B+ trees to optimize query performance, and data compression algorithms on distributed node clients reduce bandwidth consumption. The generation frequency of digital digests is synchronized with the data update rate, and the storage layout of smart contracts separates hot and cold data. Distributed ledger sharding technology improves throughput, and the returned results support multiple encoding methods. Private key backup strategies prevent data loss, and the initial runtime deviation metric set's checksum verifies the correctness of data transmission. The time window size for the time deviation component is configurable, and the dimensionality reduction method for the environmental deviation component preserves key features. The implementation of secure hash algorithms is cryptographically authenticated, and the comparison of digital digests uses a constant-time algorithm to prevent time-series attacks. The blockchain network's transaction fee model optimizes resource allocation, and smart contract gas estimation predicts execution costs. A caching layer on distributed node clients reduces blockchain access latency, and standardization processing of the initial runtime deviation metric set eliminates system differences. The weighting strategy of moving average filtering emphasizes recent data, and interpolation is used for outlier removal surrogate values. Statistical analysis of the runtime deviation metric set generates statistical reports.

[0044] Example 5: In the implementation of updating the control parameters of the target inspection robot based on the control parameter adjustment ratio, the system inputs the control parameter adjustment ratio into the fuzzy logic controller. The fuzzy logic controller internally establishes a knowledge base containing multiple fuzzy rules. The fuzzy logic controller calculates the control output based on the ratio value. The control output includes a speed adjustment and a direction adjustment. The speed adjustment is a scalar value representing the change in the robot's linear velocity, and the direction adjustment is an angle value representing the change in the robot's turning angle. The speed and direction adjustments are sent to the inspection robot's motion control unit, which is integrated into the robot's underlying controller. The actual motion trajectory is adjusted, and this adjustment is achieved by changing the difference in the robot's wheel speeds to perform the turning action.

[0045] To illustrate with a concrete example, when a power plant inspection robot performs an inspection task in a cable tunnel within a confined space, the calculated control parameter adjustment ratio is 0.75. This value is input to the fuzzy logic controller (FLC). The FLC's fuzzification interface converts the precise input value of 0.75 into a fuzzy set. The FLC's rule base contains fuzzy rules such as "if the adjustment ratio is large, the speed adjustment is positively large and the direction adjustment is positively small." The FLC's inference engine activates the matching rules, and the FLC's defuzzification interface converts the fuzzy output into a precise speed adjustment of 0.2 m / s and a direction adjustment of 15 degrees. These values ​​are sent to the motion control unit (PCU), which parses the instructions and generates PWM signals to drive the motors. The robot's actual trajectory changes from straight-line travel to a curve with a 15-degree turning angle, while the speed increases by 0.2 m / s. After constructing the operating state deviation vector, the system performs minimum-maximum scaling on each component of the vector, linearly transforming the original data to a specified range. The numerical range of all components is adjusted to between 0 and 1. This uniformity improves the stability of subsequent processing algorithms. The scaled runtime deviation vector is used in subsequent retrieval steps, which are performed on standardized data to enhance matching accuracy.

[0046] The fuzzy logic controller (FLC) is designed using a Mamdani-type fuzzy inference system. The input variables are the control parameter adjustment ratios, and the output variables include speed and direction adjustments. The universe of discourse for the control parameter adjustment ratios is divided into three fuzzy sets: "small," "medium," and "large." The universe of discourse for the speed adjustment is divided into five fuzzy sets: "negative large," "negative small," "zero," "positive small," and "positive large." The universe of discourse for the direction adjustment is divided into five fuzzy sets: "left large," "left small," "zero," "right small," and "right large." The FLC's rule base contains 3 × 5 × 5 = 75 fuzzy rules, each describing the input-output relationship in an "if-then" format. The fuzzification method uses a triangular membership function, the inference engine uses a max-min synthesis method, and the defuzzification method uses the centroid method to calculate the precise output value. The motion control unit's hardware architecture is based on a 32-bit microprocessor, and its software module implements a PID control algorithm. Speed ​​adjustments are converted into motor speed commands, and direction adjustments are converted into wheel speed difference commands. Tracking of the actual motion trajectory is achieved through encoder feedback to form a closed-loop control system, with obstacle avoidance constraints considered during trajectory planning. The robot's mechanical structure employs a four-wheel differential drive, with each drive wheel equipped with an independent DC motor and reducer. The motion control unit's communication interface uses the CAN bus protocol, and its sampling period is 10 milliseconds. The minimum-maximum scaling algorithm includes extreme value statistics, and the component values ​​of the running state deviation vector are updated in real time. The parameters for adjusting the numerical range are derived from historical data statistics, with the historical data statistics window containing the most recent 1000 sampling points. The storage format of the scaled running state deviation vector optimizes memory usage, and Euclidean distance is used for similarity calculation in subsequent retrieval steps. Standardization prevents components with larger dimensions from dominating the retrieval results, and dynamic range compression of vector components improves the algorithm's numerical stability.

[0047] In the extended scenario of the cable tunnel inspection example, when the robot detects a narrow section of the pipeline ahead, the control parameter adjustment ratio may change to 0.9. The fuzzy logic controller activates the rule subset corresponding to "large adjustment ratio," and the inference result may produce a larger speed adjustment and a more cautious direction adjustment. Simultaneously, the running state deviation vector may add a narrow space detection component, and the scaling processing of this component requires expanding the parameter table of the minimum-maximum scaling process. After receiving the adjustment command, the motion control unit controls the robot to decelerate and turn through the narrow area with a smaller radius. The parameter tuning of the fuzzy logic controller was completed experimentally, and the boundary of the fuzzy set was optimized. The calibration procedure of the motion control unit matches the motor characteristics, and the accuracy of the actual motion trajectory is verified using a laser rangefinder. The extreme value update strategy of the minimum-maximum scaling process uses a sliding window, and the dimension of the running state deviation vector can vary with the sensor configuration. The scaling process robustly handles abnormal input conditions, and the efficiency of the retrieval step is optimized through indexing techniques. The real-time performance of the system meets the requirements of the inspection task, and the computational latency of the fuzzy logic controller is less than 1 millisecond. The response time of the motion control unit ensures timely obstacle avoidance by the robot, and the computational overhead of the minimum-maximum scaling process is negligible. The entire control parameter update process is completed within a 100-millisecond cycle, adapting to the real-time requirements of the power plant environment. The robot's motion smoothness is ensured through gradual adjustments, and safety is maintained through a multi-level verification mechanism. During long-term operation, the fuzzy logic controller's rule base supports online learning and adjustment, and the motion control unit's parameters support adaptive tuning. The extreme value statistics of the minimum-maximum scaling process support dynamic updates, and the composition of the operating state deviation vector supports configuration changes. The system's scalability supports adding new control dimensions, and its compatibility supports different models of inspection robot platforms. The maintenance interface supports parameter monitoring and fault diagnosis, and the log system records the complete control decision-making process. A cable tunnel inspection example demonstrates the system's adaptability in complex environments, showcasing the advantages of the fuzzy logic controller in handling uncertainty. The precise execution of the motion control unit ensures accurate robot movements, and the standardization of the minimum-maximum scaling process enhances system robustness.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A control method for a power plant narrow space inspection robot, characterized by, The method comprises the following steps: When a preset inspection trigger condition is met, a historical running state of a target inspection robot at a specified historical time is obtained, wherein the historical running state comprises a historical cumulative running time length and a historical environment parameter; An real-time running state of the target inspection robot is obtained, wherein the real-time running state comprises a real-time cumulative running time length and a real-time environment parameter; a running state deviation vector is constructed according to the historical running state and the real-time running state; a running deviation metric set corresponding to the running state deviation vector is retrieved; when a distribution dispersion degree of the running deviation metric set is greater than or equal to a dispersion threshold, a ratio of a historical duration from a storage start time to the specified historical time to a current duration from the storage start time to a current time is calculated as a control parameter adjustment ratio; The control parameter of the target inspection robot is updated according to the control parameter adjustment ratio.

2. The control method of the power station narrow space inspection robot according to claim 1, characterized in that, The step of obtaining the historical running state of the target inspection robot at the specified historical time comprises the following steps: obtaining a plurality of environment parameter sequences of the target inspection robot; for each environment parameter sequence, a dynamic time warping distance between the environment parameter sequence and a historical reference sequence is calculated; based on the dynamic time warping distance, a heterogenic running time length corresponding to each environment parameter sequence is determined; a minimum value in all heterogenic running time lengths is extracted, and the minimum value is multiplied by a preset coefficient to obtain the preset inspection trigger condition.

3. The control method of the power station narrow space inspection robot according to claim 1, characterized in that, The step of obtaining the historical running state of the target inspection robot at the specified historical time further comprises the following steps: real-time monitoring of the running data volume of the target inspection robot; calculating an instantaneous growth rate of the running data volume; when the instantaneous growth rate exceeds a preset rate threshold, automatically triggering the obtaining of the historical running state of the target inspection robot at the specified historical time.

4. The control method of the power station narrow space inspection robot according to claim 2, characterized in that, The step of calculating, for each environment parameter sequence, a dynamic time warping distance between the environment parameter sequence and a historical reference sequence comprises the following steps: when the environment parameter sequence is an original sensor reading, a plurality of historical sensor reading sequences and corresponding running time length sequences are retrieved from a historical database; for each historical sensor reading sequence and the current sensor reading sequence, a dynamic time warping calculation is performed to obtain a warping path distance of each sequence pair; all warping path distances are collected to form a distance set; a warping path distance subset greater than a preset distance threshold is selected from the distance set; a mode calculation is performed on the running time length sequences corresponding to the warping path distance subset to obtain the heterogenic running time length.

5. The control method of the power station narrow space inspection robot according to claim 2, characterized in that, The step of calculating, for each environment parameter sequence, a dynamic time warping distance between the environment parameter sequence and a historical reference sequence comprises the following steps: when the environment parameter sequence is a fusion feature, a feature extraction model associated with the fusion feature is called, wherein the feature extraction model is constructed based on a pre-trained neural network; historical fusion feature sequences are obtained by processing historical sensor data through the feature extraction model; a cosine similarity between the historical fusion feature sequences and a current fusion feature sequence is calculated; according to the deviation of the cosine similarity from a preset similarity threshold, the heterogenic running time length is derived.

6. The control method of the power station narrow space inspection robot according to claim 5, characterized in that, The method further comprises the following steps: The network structure of the feature extraction model is configured, and the parameters of the convolutional layer and the fully connected layer are specified; Collect multiple sets of training data, each set of training data including a sequence of sensor readings and a corresponding fusion feature label; use a gradient descent algorithm to optimize the weights of the feature extraction model; bind the optimized feature extraction model with a specific fusion feature and environmental parameters.

7. The control method of the power station narrow space inspection robot according to claim 1, characterized in that, The step of searching for a set of operation deviation metrics corresponding to the operation state deviation vector includes: decomposing the operation state deviation vector into a time deviation component and an environmental deviation component; encrypting the time deviation component and the environmental deviation component using a secure hash algorithm to generate a digital digest; sending the digital digest to a blockchain network to retrieve a matching operation deviation record through a deployed smart contract; obtaining an initial set of operation deviation metrics from the returned result of the smart contract; performing moving average filtering processing on the initial set of operation deviation metrics to remove outliers to obtain the set of operation deviation metrics.

8. The control method of claim 7, wherein the robot is a robot for inspecting a narrow space of a power plant. The step of sending the digital digest to the blockchain network includes: deploying a retrieval smart contract in the blockchain network, which contains matching logic based on deviation components; calling the smart contract through a distributed node client and inputting the digital digest as an input parameter; the smart contract queries records that meet the deviation condition in the distributed ledger and returns encrypted query results; decrypt the query results using a private key to obtain the initial set of operation deviation metrics.

9. The control method of a power plant narrow space inspection robot according to claim 1, wherein The step of updating the control parameters of the target inspection robot according to the control parameter adjustment scale includes: inputting the control parameter adjustment scale into a fuzzy logic controller; the fuzzy logic controller calculates the control output according to the scale value, including the speed adjustment amount and the direction adjustment amount; send the speed adjustment amount and the direction adjustment amount to the motion control unit of the inspection robot to adjust the actual motion trajectory.

10. The control method of the power station narrow space inspection robot according to claim 1, characterized in that, Also includes: After constructing the operation state deviation vector, perform min-max scaling on each component of the operation state deviation vector to adjust the numerical range of all components to between 0 and 1; use the scaled operation state deviation vector for subsequent retrieval steps.

Images