Excavator electric pole square hole foundation pit excavation posture real-time correction control method and system based on deep learning

Abstract

The invention discloses an excavator electric pole square hole foundation pit excavation attitude real-time correction control method and system based on deep learning, relates to the field of square hole foundation pit excavation, and guarantees data reliability through multi-source sensor synchronous acquisition and CAN bus + Ethernet double-link redundancy transmission. After preprocessing standardization, inputting a CNN-LSTM mixed model which is subjected to transfer learning training and quantitative optimization to realize high-precision attitude recognition and multi-dimensional deviation calculation, and generating a targeted correction instruction with clear priority by combining a dynamically adaptive reference attitude parameter library and two-dimensional fuzzy PID control; then a safe closed-loop control system is constructed through closed-loop feedback incremental fine adjustment and three-level anomaly grading treatment, real-time accurate correction of the excavation posture of the electric pole square hole foundation pit can be achieved without manual intervention, the consistency of the geometric size, perpendicularity and depth of the foundation pit is improved, the method can also adapt to different soil types, excavation depths and working environments, and the safety of the electric pole square hole foundation pit is improved. And the labor intensity and manual operation errors are greatly reduced.

Description

Technical Field

[0001] This invention relates to square hole foundation pit excavation technology, specifically to a method and system for real-time correction and control of the excavator pole square hole foundation pit excavation posture based on deep learning. Background Technology

[0002] With the continuous advancement of power grid infrastructure construction and the rapid development of smart grids in my country, the laying and maintenance of power lines have placed higher demands on construction efficiency and project quality. Taking Kashgar region as an example, in the past five years, rural power grid renovation projects have required the erection of at least 25,000 power poles annually. Among these, special pole types such as corner poles, terminal poles, equipment installation poles for transformers and circuit breakers, and straight poles in soft sand dune areas, which require the installation of square cement base plates, require the excavation of more than 10,000 square hole foundation pits. Although the individual project volume of such foundation pits is small, the number is huge. Traditional manual excavation is inefficient and of poor quality, making it difficult to meet the basic requirements of project progress and quality control. As a fundamental link in the entire power project, the excavation accuracy of the square hole foundation pits for power poles is directly related to the verticality, stability, and safe operation of subsequent power lines.

[0003] Currently, the excavation of square hole foundation pits for utility poles mainly relies on two methods: traditional manual excavation and semi-automatic excavator excavation. Traditional manual excavation depends on the experience of the operators and is carried out by excavating with hand tools or small machinery. This not only results in high labor intensity and extremely low work efficiency, but also makes it difficult to ensure the consistency of the geometric dimensions, verticality, and depth of the square holes, thus affecting the subsequent processes and the overall safety of the project. Although semi-automatic excavator excavation improves efficiency, the degree of automation is still relatively low. It still requires skilled operators to control the bucket posture and trajectory through visual observation and experience, making it difficult to achieve high-precision excavation. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time correction and control method and system for the excavation posture of excavators for square hole foundation pits on utility poles based on deep learning, so as to solve the problem that although semi-automatic excavators in the prior art improve efficiency, their degree of automation is insufficient, and they still need manual judgment and control, making it difficult to achieve high-precision excavation.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a real-time correction and control method for the excavation posture of a square hole foundation pit for an excavator based on deep learning, comprising the following steps:

[0006] Step S1: Acquire real-time multi-source data; Using the attitude sensor, vision sensor, and laser rangefinder integrated into the excavator's electronic control system, a unified synchronous triggering mechanism is used to synchronously collect the excavator bucket attitude parameters, pit outline data, and relative position data of the bucket and square hole reference trajectory during the excavation process. The acquisition frequency is not less than 100Hz and can be adaptively adjusted according to the excavation depth; The raw data stream is formed based on CAN bus + Ethernet dual-link redundant transmission, and a hierarchical storage strategy is adopted to ensure data reliability.

[0007] Step S2: Data preprocessing and standardization; outlier removal, noise filtering and data alignment are performed on the raw data stream. The attitude parameters and position data are uniformly mapped to the preset three-dimensional rectangular coordinate system of the square hole foundation pit to obtain a standardized data matrix. The preprocessing adopts an adaptive Kalman filter algorithm based on a sliding window. The window size is dynamically adjusted according to the excavation depth.

[0008] Step S3: Deep learning model-driven attitude recognition and deviation calculation; Input the standardized data matrix into the CNN-LSTM hybrid model trained and quantized by transfer learning, output the current excavation attitude parameters of the bucket through multi-source data fusion, compare it with the dynamically adapted square hole foundation pit excavation benchmark attitude parameters, and obtain multi-dimensional attitude deviation values, including angle deviation, position deviation, depth deviation and comprehensive deviation.

[0009] Step S4: Generation of correction control commands; Based on the priority sorting of attitude deviation values, targeted attitude correction control commands are generated through a two-dimensional fuzzy PID control algorithm. The control commands include bucket slewing angle correction commands, stick extension length correction commands, and boom lifting height correction commands, which are output after being verified by the excavator kinematic model.

[0010] Step S5: Real-time execution and feedback optimization; The correction control command is sent to the excavator's electronic control execution unit, the command execution status is monitored in real time, the corrected posture data is collected and simplified preprocessed to generate a feedback data matrix, the adjustment accuracy trigger model is incrementally fine-tuned, and an anomaly classification handling mechanism is adopted to ensure operational safety, thereby realizing closed-loop real-time correction control of the excavation posture.

[0011] Further, step S1 includes the following steps:

[0012] Step S11: Obtain the excavator bucket attitude parameters, pit outline data, and relative position data between the bucket and the square hole reference trajectory;

[0013] Step S12: Use CAN bus to transmit attitude and laser data, and Ethernet to transmit vision data. Each link has a dedicated channel. The multi-sensor protocol is uniformly converted to CANopen protocol through the built-in protocol conversion module of the PLC controller.

[0014] Step S13: Generate a unified synchronous trigger signal through the PLC controller, and the acquisition frequency is adaptively adjusted according to the excavation depth and operation status.

[0015] Further, step S2 includes the following steps:

[0016] Step S21: Identify outliers and obtain outlier data;

[0017] Step S22: Perform targeted noise filtering on attitude parameters, laser distance data, and visual image data;

[0018] Step S23: Perform time alignment on the multi-source data to obtain aligned data;

[0019] Step S24: Construct a three-dimensional rectangular coordinate system with the geometric center of the square hole foundation pit as the origin, and map the aligned data to this coordinate system to obtain a standardized data matrix.

[0020] Furthermore, step S3 includes the following steps:

[0021] Step S31: Construct a million-level dataset covering different soil types, excavation depths and working environments, perform data augmentation, train a CNN-LSTM hybrid model using transfer learning strategy, load pre-trained weights through ResNet50, and perform quantization processing through TensorRTINT8;

[0022] Step S32: Integrate multi-sensor data to obtain attitude parameters, cross-validate them with laser and vision data, and determine the parameters to be valid if the verification error is ≤1mm;

[0023] Step S33: Based on the design dimensions of the square hole foundation pit, soil type, and detailed benchmark posture parameters of the excavation stage, construct a local SQLite parameter library, and dynamically adjust the benchmark parameters based on the actual excavation contour during operation;

[0024] Step S34: Calculate the angle deviation, position deviation, depth deviation and comprehensive deviation respectively; filter the instantaneous deviation and integrate them into a three-dimensional deviation vector output with confidence label.

[0025] Further, step S4 includes the following steps:

[0026] Step S41: Construct a two-dimensional fuzzy controller with the inputs being the deviation value e and the deviation change rate ec, and the outputs being the PID parameter corrections ΔKp, ΔKi, and ΔKd. Formulate a 7×7 fuzzy control rule table.

[0027] Step S42: Obtain the bucket slewing angle correction command based on the angle deviation, obtain the stick extension length correction command based on the position deviation, and obtain the boom lifting height correction command based on the depth deviation and pitch angle deviation;

[0028] Step S43: Instruction optimization and verification; sort instructions according to priority order of angle deviation > depth deviation > position deviation, verify the validity of instructions, convert them into a format recognizable by the electronic control unit and add a check code, with instruction transmission delay ≤ 5ms.

[0029] Further, step S5 includes the following steps:

[0030] Step S51: Send commands according to priority via CAN bus, calculate the coordination parameters of the actuator when multiple commands are executed in coordination, monitor the execution status and progress of the action in real time, and trigger a reset and alarm when a jam or deviation occurs.

[0031] Step S52: After the correction command is executed, delay for 50ms, collect feedback data at a frequency of 200Hz for 1s, and after outlier removal, data alignment and coordinate mapping simplification preprocessing, take the average value to obtain the feedback data matrix and calculate the adjustment accuracy;

[0032] Step S53: Trigger incremental fine-tuning based on the results of the precision grading, only adjust the weights of the fully connected layer and LSTM layer of the model, select the feedback data of the most recent 1000 frames as the fine-tuning dataset, and verify it through the test dataset after fine-tuning.

[0033] Step S54: Locate the source of the anomaly through a three-level anomaly classification mechanism, and execute prompts, audible and visual alarms, and emergency switching to manual control mode according to the anomaly level. After the anomaly is resolved, automatically switch back to automatic control mode.

[0034] Further, step S33 includes the following steps:

[0035] Step S331: Benchmark parameter subdivision and adaptation; subdivide the benchmark parameters according to every 0.5m excavation depth and the initial / middle / final stages of excavation, and preset corresponding parameters for square hole side lengths of 0.5-1.5m and depths of 2-5m respectively;

[0036] Step S332: Dynamic adjustment of reference parameters; Based on the actual contour of the foundation pit identified by the visual sensor, when the excavation deviation of a certain side wall is >10cm, the reference posture parameters of the corresponding side excavation are automatically adjusted, with an adjustment range ≤0.5°.

[0037] Further, step S53 includes the following steps:

[0038] Step S531: Fine-tuning trigger conditions; model fine-tuning is triggered when the adjustment accuracy is <70%, when there are a total of 5 second-level accuracy samples, or when the working environment changes.

[0039] Step S532: Fine-tune parameter constraints; the adjustment range of the weights of the fully connected layer is ≤ ±5%, the adjustment range of the weights of the LSTM layer is ≤ ±3%, the fine-tuning learning rate is set to 0.0001, the fine-tuning rounds are 10, and the fine-tuning time is ≤ 5 min.

[0040] Further, step S54 includes the following steps:

[0041] Step S541: Anomaly classification criteria; Level 1 anomaly is single frame data loss, and Level 2 adjustment accuracy is also anomaly; Level 2 anomaly is sensor failure, and Level 3 adjustment accuracy is also anomaly; Level 3 anomaly is actuator failure, and Model inference error.

[0042] Step S542: Graded processing flow; Level 1 anomalies are automatically compensated, Level 2 anomalies trigger audible and visual alarms and emergency measures, Level 3 anomalies trigger emergency alarms and switch to manual control mode, while locking the actuator.

[0043] Furthermore, the present invention also provides a deep learning-based real-time posture correction and control system for excavating square hole foundation pits on excavators, used to execute the deep learning-based real-time posture correction and control method for excavating square hole foundation pits on excavators as described above. The system includes:

[0044] The multi-source data acquisition module is used to synchronously acquire bucket attitude parameters, pit contour data and relative position data through redundantly deployed attitude sensors, vision sensors and laser rangefinders, using a synchronous triggering mechanism and dual-link redundant transmission, forming a raw data stream with a check code, and supporting adaptive adjustment of acquisition frequency and hierarchical storage.

[0045] The data preprocessing and standardization module is used to perform outlier removal, adaptive Kalman filter noise filtering, and spatiotemporal alignment on the raw data stream, construct a three-dimensional rectangular coordinate system for the square hole foundation pit, and uniformly map and standardize the data into a standardized data matrix in the [0,1] interval;

[0046] The pose recognition and deviation calculation module is used to load the CNN-LSTM hybrid model trained and optimized by transfer learning, output the current pose parameters through multi-source data fusion, call the dynamically adapted baseline pose parameter library, calculate multi-dimensional pose deviation vectors and optimize filtering.

[0047] The correction control command generation module is used to obtain targeted correction control commands based on deviation priority sorting through a two-dimensional fuzzy PID controller. After verification and format conversion by the excavator kinematic model, the commands are sent to the electronic control execution unit according to priority.

[0048] The execution feedback optimization module is used to monitor the command execution status in real time, collect and preprocess feedback data, calculate and adjust accuracy and trigger incremental fine-tuning of the model. It adopts a three-level anomaly classification and processing mechanism to realize fault location, alarm and mode switching, and build a closed-loop correction control system.

[0049] Compared with existing technologies, this invention provides a real-time correction and control method and system for the excavation posture of a square hole foundation pit for a utility pole based on deep learning. It ensures data reliability through synchronous acquisition from multiple sources of sensors and redundant transmission via CAN bus + Ethernet dual links. After preprocessing and standardization, the data is input into a CNN-LSTM hybrid model trained and quantized for optimization, achieving high-precision posture recognition and multi-dimensional deviation calculation. Combined with a dynamically adaptable reference posture parameter library and two-dimensional fuzzy PID control, it generates targeted correction commands with clear priorities. Furthermore, it constructs a safe closed-loop control system through closed-loop feedback incremental fine-tuning and three-level anomaly classification processing. This allows for real-time and accurate correction of the excavation posture of the square hole foundation pit for the utility pole without manual intervention. It not only controls the deviation verification error to ≤0.5mm and the command transmission delay to ≤5ms, improving the consistency of the foundation pit's geometric dimensions, verticality, and depth, but also adapts to different soil types, excavation depths, and operating environments, significantly reducing labor intensity and human error. This provides core technical support for the subsequent stability of pole installation and the overall safety of power engineering. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0051] Figure 1 This is a schematic diagram of the steps provided in an embodiment of the present invention;

[0052] Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S1.

[0053] Figure 3 for Figure 1 A detailed flowchart illustrating the implementation steps of step S2.

[0054] Figure 4 for Figure 1 A detailed flowchart illustrating the implementation steps of step S3.

[0055] Figure 5 for Figure 1 A detailed flowchart illustrating the implementation steps of step S4.

[0056] Figure 6 for Figure 1 A detailed flowchart illustrating the implementation steps of step S5. Detailed Implementation

[0057] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0058] To address the issue that while semi-automated excavators in existing technologies improve efficiency, their insufficient automation, reliance on manual judgment and control, and difficulty in achieving high-precision excavation, please refer to [the relevant documentation / reference]. Figure 1 - Figure 6 The following preferred technical solutions are provided.

[0059] One embodiment of the present invention provides a real-time correction and control method for the excavation posture of an excavator in a square hole foundation pit for utility poles, based on deep learning, comprising the following steps:

[0060] Step S1: Acquire real-time multi-source data; Using the attitude sensor, vision sensor and laser rangefinder integrated into the excavator's electronic control system, a unified synchronous triggering mechanism is used to synchronously collect the excavator bucket attitude parameters, pit outline data and relative position data of the bucket and square hole reference trajectory during the excavation process; The raw data stream is formed based on CAN bus + Ethernet dual-link redundant transmission;

[0061] Step S2: Data preprocessing and standardization; outlier removal, noise filtering and data alignment are performed on the raw data stream, and the attitude parameters and position data are uniformly mapped to the preset three-dimensional rectangular coordinate system of the square hole foundation pit to obtain a standardized data matrix.

[0062] Step S3: Deep learning model-driven attitude recognition and deviation calculation; Input the standardized data matrix into the CNN-LSTM hybrid model trained by transfer learning and quantized and optimized, output the current excavation attitude parameters of the bucket through multi-source data fusion, and compare it with the dynamically adapted square hole foundation pit excavation benchmark attitude parameters to obtain multi-dimensional attitude deviation values.

[0063] Step S4: Generation of correction control commands; Based on the priority ranking of attitude deviation values, targeted attitude correction control commands are generated using a two-dimensional fuzzy PID control algorithm;

[0064] Step S5: Real-time execution and feedback optimization; The correction control command is sent to the excavator's electronic control execution unit, the command execution status is monitored in real time, the corrected posture data is collected and simplified preprocessed to generate a feedback data matrix, the adjustment accuracy triggers the model for incremental fine-tuning, and an anomaly classification handling mechanism is adopted to ensure operational safety.

[0065] In this embodiment of the invention, reference Figure 1The above is a flowchart illustrating the steps of the present invention. In this example, a real-time correction and control method for the excavation posture of a square hole foundation pit for an excavator based on deep learning includes the following steps:

[0066] Step S1: Acquire real-time multi-source data; Using the attitude sensor, vision sensor and laser rangefinder integrated into the excavator's electronic control system, a unified synchronous triggering mechanism is used to synchronously collect the excavator bucket attitude parameters, pit outline data and relative position data of the bucket and square hole reference trajectory during the excavation process; The raw data stream is formed based on CAN bus + Ethernet dual-link redundant transmission;

[0067] In this embodiment of the invention, the industrial-grade temperature- and dust-resistant IMU inertial measurement unit, high-definition vision sensor, and pulsed laser rangefinder integrated into the excavator's electronic control system are used to achieve synchronous acquisition of multi-source data using a unified synchronous trigger signal generated by the PLC controller. The acquisition frequency is not less than 100Hz and can be adaptively adjusted according to the excavation depth and operating status. Attitude and laser data are transmitted via CAN bus, and vision data is transmitted via Ethernet. Each link has a dedicated channel, and the protocol conversion module unifies them into the CANopen protocol to form a raw data stream containing a checksum. At the same time, a hierarchical storage strategy of local SSD solid-state drive + cloud OSS object storage is used to ensure data reliability.

[0068] Step S2: Data preprocessing and standardization; outlier removal, noise filtering and data alignment are performed on the raw data stream, and the attitude parameters and position data are uniformly mapped to the preset three-dimensional rectangular coordinate system of the square hole foundation pit to obtain a standardized data matrix.

[0069] In this embodiment of the invention, outliers in the original data stream are removed by combining the 3σ criterion with the isolated forest algorithm. Noise is filtered by an adaptive Kalman filter algorithm. The window size is dynamically adjusted with the excavation depth. The spatiotemporal alignment of multi-source data is completed based on the attitude sensor timestamp, with an error ≤1ms. A three-dimensional rectangular coordinate system is constructed with the geometric center of the square hole foundation pit as the origin. The min-max normalization method is used to map the aligned data to this coordinate system to generate a normalized data matrix in the [0,1] interval, providing a unified data foundation for subsequent model calculations.

[0070] Step S3: Deep learning model-driven attitude recognition and deviation calculation; Input the standardized data matrix into the CNN-LSTM hybrid model trained by transfer learning and quantized and optimized, output the current excavation attitude parameters of the bucket through multi-source data fusion, and compare it with the dynamically adapted square hole foundation pit excavation benchmark attitude parameters to obtain multi-dimensional attitude deviation values.

[0071] In this embodiment of the invention, a dataset of millions of data points covering different soil types, excavation depths, and working environments is constructed and augmented. A CNN-LSTM hybrid model is trained using a transfer learning strategy, and a ResNet50 model is loaded with pre-trained weights. After TensorRTINT8 quantization, the model is deployed on an edge computing terminal. The current pose parameters are obtained by fusing multi-sensor data through Kalman filtering and cross-validated with laser and visual data. An error ≤0.5mm is considered valid. The local SQLite benchmark parameter library is called, and the benchmark parameters are dynamically adjusted based on the actual pit outline, according to the excavation depth, stage, etc. The angle, position, depth, and comprehensive deviation are calculated. A confidence level ≥95% is considered valid, thus completing pose recognition and deviation quantification.

[0072] Step S4: Generation of correction control commands; Based on the priority ranking of attitude deviation values, targeted attitude correction control commands are generated using a two-dimensional fuzzy PID control algorithm;

[0073] In this embodiment of the invention, a two-dimensional fuzzy PID controller is constructed. Taking the deviation value e and the deviation change rate ec as inputs, the controller outputs PID parameter correction quantities ΔKp, ΔKi, and ΔKd. A 7×7 fuzzy control rule table is formulated and clarified using the center of gravity method. Based on the angle, position, and depth deviations, correction commands for bucket rotation angle, stick extension length, and boom lifting height are generated respectively. These commands are prioritized according to the order of angle deviation > depth deviation > position deviation. After verification by the excavator's kinematic model, the commands are converted into a format recognizable by the electronic control unit (with added checksums) to ensure that the command transmission delay is ≤5ms, achieving targeted correction.

[0074] Step S5: Real-time execution and feedback optimization; Send the correction control command to the excavator's electronic control execution unit, monitor the command execution status in real time, collect the corrected posture data, generate a feedback data matrix after simplified preprocessing, and incrementally fine-tune the model based on the adjustment accuracy trigger. At the same time, an anomaly classification handling mechanism is adopted to ensure operational safety.

[0075] In this embodiment of the invention, correction control commands are issued according to priority via the CAN bus. When multiple commands are executed in coordination, the coordination parameters of the actuator are calculated, and the execution status of the action is monitored in real time. A reset alarm is triggered when there is a lag or out-of-tolerance. After the command is executed, there is a 50ms delay, and feedback data is collected at a frequency of 200Hz for 1 second. After simplified preprocessing, a feedback data matrix is ​​generated and the adjustment accuracy is calculated. Based on the adjustment accuracy, incremental fine-tuning of the model is triggered, adjusting only the weights of specific layers. At the same time, a three-level anomaly classification and processing mechanism is adopted to locate the anomaly source and execute corresponding measures, thus constructing a closed-loop real-time correction control system to ensure operational accuracy and safety.

[0076] Step S1 includes the following steps:

[0077] Step S11: Obtain the excavator bucket attitude parameters, pit outline data, and relative position data between the bucket and the square hole reference trajectory;

[0078] Step S12: Use CAN bus to transmit attitude and laser data, and Ethernet to transmit vision data. Each link has a dedicated channel. The multi-sensor protocol is uniformly converted to CANopen protocol through the built-in protocol conversion module of the PLC controller.

[0079] Step S13: Generate a unified synchronous trigger signal through the PLC controller, and the acquisition frequency is adaptively adjusted according to the excavation depth and operation status.

[0080] As an example of the present invention, reference is made to... Figure 2 As shown, step S1 in this example includes:

[0081] Step S11: Obtain the excavator bucket attitude parameters, pit outline data, and relative position data between the bucket and the square hole reference trajectory;

[0082] In this embodiment of the invention, the acquired parameters include attitude parameters such as the bucket's angle and pitch angle, contour data such as the pit's outline dimensions and shape, and relative position data such as the offset between the bucket's actual trajectory and the square hole reference trajectory. Sensor selection and deployment must meet industrial-grade operational requirements: the IMU (Inertial Measurement Unit) is deployed at the bucket boom connection and the boom root, using a titanium alloy bracket welded and bolted, and equipped with shock-absorbing pads; an industrial-grade high-definition camera, equipped with an anti-fog and anti-glare lens and an infrared supplementary light, is deployed on the top of the cab and the side of the bucket; a pulsed laser rangefinder is deployed at the tip of the bucket. The boom top and stick middle are redundantly configured, and the launch port is equipped with a dust cover. At the same time, sensor calibration and environmental adaptation are required, including static and dynamic dual-mode calibration of the IMU, Zhang Zhengyou calibration of the vision sensor, gain adjustment of the laser rangefinder for different soil types (1.2 times for soft soil, 0.8 times for rock), and automatic switching of the backup sensor when the main sensor packet loss rate exceeds 1%, with a switching delay of ≤10ms. The static and dynamic dual-mode calibration of the IMU is repeated every 50 hours, and the Zhang Zhengyou calibration of the vision sensor is repeated every 100 hours.

[0083] Step S12: Use CAN bus to transmit attitude and laser data, and Ethernet to transmit vision data. Each link has a dedicated channel. The multi-sensor protocol is uniformly converted to CANopen protocol through the built-in protocol conversion module of the PLC controller.

[0084] In this embodiment of the invention, the division of labor between the two links is clearly defined: the CAN bus is responsible for transmitting data from the attitude sensor and the laser rangefinder, while the Ethernet is responsible for transmitting image data from the vision sensor. Both links are equipped with backup channels to ensure transmission redundancy. The different protocols of the attitude sensor, vision sensor, and laser rangefinder are uniformly converted to the CANopen protocol through the protocol conversion module built into the PLC controller to ensure data interaction compatibility. After the protocol conversion is completed, a 100-hour continuous joint debugging test is required to verify the stability of the link transmission and the consistency of the data.

[0085] Step S13: Generate a unified synchronous trigger signal through the PLC controller, and the acquisition frequency is adaptively adjusted according to the excavation depth and operation status;

[0086] In this embodiment of the invention, a unified synchronous trigger signal is generated by the PLC controller, and the trigger frequency is adaptively adjusted according to the excavation depth and operation status; the collected single-frame data is encapsulated into JSON format with CRC32 check code according to the timestamp to avoid tampering or loss during data transmission; the storage adopts a hierarchical strategy, the local SSD solid-state drive is automatically cleaned up after 7 days, and the cloud OSS object storage is retained for a long time, balancing storage efficiency and data security.

[0087] Step S2 includes the following steps:

[0088] Step S21: Identify outliers and obtain outlier data;

[0089] Step S22: Perform targeted noise filtering on attitude parameters, laser distance data, and visual image data;

[0090] Step S23: Perform time alignment on the multi-source data to obtain aligned data;

[0091] Step S24: Construct a three-dimensional rectangular coordinate system with the geometric center of the square hole foundation pit as the origin, and map the aligned data to this coordinate system to obtain a standardized data matrix.

[0092] As an example of the present invention, reference is made to... Figure 3 As shown, step S2 in this example includes:

[0093] Step S21: Identify outliers and obtain outlier data;

[0094] In this embodiment of the invention, outliers are identified by combining the 3σ criterion with the isolated forest algorithm. The 3σ criterion is used to filter out data outliers that conform to a normal distribution, while the isolated forest algorithm is used to detect isolated outliers that do not conform to a normal distribution. For the identified outliers, linear interpolation is used to replace them to ensure the continuity of the data sequence and avoid outliers from interfering with subsequent preprocessing and model calculations.

[0095] Step S22: Perform targeted noise filtering on attitude parameters, laser distance data, and visual image data;

[0096] In this embodiment of the invention, an adaptive Kalman filter algorithm based on a sliding window is used for targeted noise filtering: differentiated filtering parameters are set for different noise characteristics of attitude parameters, laser distance data, and visual image data; the size of the sliding window is dynamically adjusted according to the excavation depth, with 5 frames when the excavation depth is ≤2m, 8 frames when it is 2-4m, and 12 frames when it is ≥4m, so as to achieve accurate noise filtering under different working scenarios.

[0097] Step S23: Perform time alignment on the multi-source data to obtain aligned data;

[0098] In this embodiment of the invention, the timestamp of the attitude sensor is used as a unified benchmark to perform time alignment on attitude data, laser data, and visual data. For visual data, Zhang Zhengyou calibration method is used to complete spatial coordinate transformation to ensure the consistency of multi-source data in time and space dimensions, and finally achieve a time alignment error of ≤1ms, ensuring the accuracy of data fusion.

[0099] Step S24: Construct a three-dimensional rectangular coordinate system with the geometric center of the square hole foundation pit as the origin, and map the aligned data to this coordinate system to obtain a standardized data matrix;

[0100] In this embodiment of the invention, a three-dimensional rectangular coordinate system is constructed, with the geometric center of the square hole foundation pit as the origin. The x-axis and y-axis are set to be perpendicular to each other along the horizontal plane of the foundation pit, and the z-axis is set along the depth direction of the foundation pit. The aligned data after outlier removal, noise filtering, and spatiotemporal alignment are uniformly mapped to this three-dimensional rectangular coordinate system. The min-max normalization method is used to process the mapped data, normalizing all data to the [0,1] interval to form a standardized data matrix and eliminate the dimensional differences of data in different dimensions.

[0101] Step S3 includes the following steps:

[0102] Step S31: Construct a million-level dataset covering different soil types, excavation depths and working environments, perform data augmentation, train a CNN-LSTM hybrid model using transfer learning strategy, load pre-trained weights through ResNet50, and perform quantization processing through TensorRTINT8;

[0103] Step S32: Integrate multi-sensor data to obtain attitude parameters, cross-validate them with laser and vision data, and determine the parameters to be valid if the verification error is ≤1mm;

[0104] Step S33: Based on the design dimensions of the square hole foundation pit, soil type, and detailed benchmark posture parameters of the excavation stage, construct a local SQLite parameter library, and dynamically adjust the benchmark parameters based on the actual excavation contour during operation;

[0105] Step S34: Calculate the angle deviation, position deviation, depth deviation and comprehensive deviation respectively; filter the instantaneous deviation and integrate them into a three-dimensional deviation vector output with confidence label.

[0106] As an example of the present invention, reference is made to... Figure 4 As shown, step S3 in this example includes;

[0107] Step S31: Construct a million-level dataset covering different soil types, excavation depths and working environments, perform data augmentation, train a CNN-LSTM hybrid model using transfer learning strategy, load pre-trained weights through ResNet50, and perform quantization processing through TensorRTINT8;

[0108] In this embodiment of the invention, the dataset covers different soil types such as clay, sand, and rock; different excavation specifications with side lengths of 0.5-1.5m and depths of 2-5m; and different working environments such as sunny days, rainy days, and foggy days. Data augmentation is performed through methods such as rotation, scaling, and noise addition. A transfer learning strategy is used to train a CNN-LSTM hybrid model, with pre-trained weights loaded into the ResNet50 part. An early stopping strategy is combined to avoid model overfitting. After the model is trained, it is quantized using TensorRTINT8 to improve inference speed. Finally, it is deployed on an industrial-grade edge computing terminal to meet real-time control requirements.

[0109] Step S32: Integrate multi-sensor data to obtain attitude parameters, cross-validate them with laser and vision data, and determine the parameters to be valid if the verification error is ≤1mm;

[0110] In this embodiment of the invention, a Kalman filter fusion algorithm is used to integrate multi-source data from attitude sensors, vision sensors, and laser rangefinders. After smoothing through a sliding window, the current excavation attitude parameters of the bucket are output. The output attitude parameters are cross-validated with the laser rangefinder data and the visual image data. When the verification error of the three is ≤1mm, the attitude parameters are determined to be valid, ensuring the accuracy and reliability of the parameters.

[0111] Step S33: Based on the design dimensions of the square hole foundation pit, soil type, and detailed benchmark posture parameters of the excavation stage, construct a local SQLite parameter library, and dynamically adjust the benchmark parameters based on the actual excavation contour during operation;

[0112] In this embodiment of the invention, the reference posture parameters are subdivided into 0.5m excavation depths according to the design dimensions of the square hole foundation pit, soil type and excavation stage, and an SQLite local parameter library is constructed. During the operation, based on the actual outline of the foundation pit identified by the visual sensor, when the excavation deviation of a certain side wall is >10cm, the excavation reference posture parameters of the corresponding side are automatically adjusted with an adjustment range ≤0.5° to achieve dynamic adaptation of the reference parameters.

[0113] Step S34: Calculate the angle deviation, position deviation, depth deviation, and overall deviation respectively; filter the instantaneous deviation and integrate them into a three-dimensional deviation vector output with a confidence level label;

[0114] In this embodiment of the invention, multi-dimensional attitude deviations are calculated respectively: the angle deviation calculation accuracy is ≤0.01°, the position deviation calculation accuracy is ≤0.1mm, the depth deviation is calculated according to the difference between the actual excavation depth and the design depth, and the comprehensive deviation is calculated by weighted summation of 0.4Δα+0.3Δβ+0.2Δγ+0.1(Δx+Δy+Δz); a 5-frame sliding window is used to filter instantaneous deviations to avoid the influence of random errors; the final deviation is integrated into a three-dimensional deviation vector output containing timestamps and credibility identifiers, and a deviation is determined to be valid when the credibility is ≥95%.

[0115] Step S4 includes the following steps:

[0116] Step S41: Construct a two-dimensional fuzzy controller with the inputs being the deviation value e and the deviation change rate ec, and the outputs being the PID parameter corrections ΔKp, ΔKi, and ΔKd. Formulate a 7×7 fuzzy control rule table.

[0117] Step S42: Obtain the bucket slewing angle correction command based on the angle deviation, obtain the stick extension length correction command based on the position deviation, and obtain the boom lifting height correction command based on the depth deviation and pitch angle deviation;

[0118] Step S43: Instruction optimization and verification; sort instructions according to priority order of angle deviation > depth deviation > position deviation, verify the validity of instructions, convert them into a format recognizable by the electronic control unit and add a check code, with instruction transmission delay ≤ 5ms.

[0119] As an example of the present invention, reference is made to... Figure 5 As shown, step S4 in this example includes;

[0120] Step S41: Construct a two-dimensional fuzzy controller with the inputs being the deviation value e and the deviation change rate ec, and the outputs being the PID parameter corrections ΔKp, ΔKi, and ΔKd. Formulate a 7×7 fuzzy control rule table.

[0121] In this embodiment of the invention, a two-dimensional fuzzy controller is constructed, with the input quantities defined as the attitude deviation value e and the deviation change rate ec, and the output quantities as the parameter correction quantities of the PID controller, namely the proportional coefficient correction quantity ΔKp, the integral coefficient correction quantity ΔKi, and the derivative coefficient correction quantity ΔKd. At the same time, a 7×7 fuzzy control rule table is formulated, covering 7 fuzzy subsets of e and ec, such as the parameter correction logic corresponding to negative large, negative medium, negative small, zero, positive small, positive medium, and positive large. The centroid method is used to clarify the fuzzy output quantities, converting the fuzzy quantities into precise PID parameter correction values.

[0122] Step S42: Obtain the bucket slewing angle correction command based on the angle deviation, obtain the stick extension length correction command based on the position deviation, and obtain the boom lifting height correction command based on the depth deviation and pitch angle deviation;

[0123] In this embodiment of the invention, correction instructions are generated based on the correspondence between multi-dimensional deviations and excavator actuators: the slewing angle that the bucket needs to adjust is calculated based on the angle deviation, and a bucket slewing angle correction instruction is generated; the extension length of the boom that needs to be extended is calculated based on the position deviation, and a boom extension length correction instruction is generated; the boom lifting height correction instruction is generated by combining the depth deviation and the pitch angle deviation, ensuring that there is a corresponding actuator adjustment instruction for each deviation dimension.

[0124] Step S43: Instruction optimization and verification; sort instructions according to priority of angle deviation > depth deviation > position deviation, verify the validity of instructions, convert them into a format that can be recognized by the electronic control unit and add a check code, with instruction transmission delay ≤5ms;

[0125] In this embodiment of the invention, the generated correction commands are sorted according to the priority of angle deviation > depth deviation > position deviation, and the commands that have a greater impact on excavation accuracy are executed first; the validity of the commands is verified by the kinematic model of the excavator to avoid command conflicts or exceeding the movement range of the actuator; the verified commands are converted into a format that can be recognized by the electronic control unit and a check code is added to prevent transmission errors; the command transmission link is optimized to ensure that the command transmission delay is ≤5ms to meet the real-time correction requirements.

[0126] Step S5 includes the following steps:

[0127] Step S51: Send commands according to priority via CAN bus, calculate the coordination parameters of the actuator when multiple commands are executed in coordination, monitor the execution status and progress of the action in real time, and trigger a reset and alarm when a jam or deviation occurs.

[0128] Step S52: After the correction command is executed, delay for 50ms, collect feedback data at a frequency of 200Hz for 1s, and after outlier removal, data alignment and coordinate mapping simplification preprocessing, take the average value to obtain the feedback data matrix and calculate the adjustment accuracy;

[0129] Step S53: Trigger incremental fine-tuning based on the results of the precision grading, only adjust the weights of the fully connected layer and LSTM layer of the model, select the feedback data of the most recent 1000 frames as the fine-tuning dataset, and verify it through the test dataset after fine-tuning.

[0130] Step S54: Locate the source of the anomaly through a three-level anomaly classification mechanism, and execute prompts, audible and visual alarms, and emergency switching to manual control mode according to the anomaly level. After the anomaly is resolved, automatically switch back to automatic control mode.

[0131] As an example of the present invention, reference is made to... Figure 6 As shown, step S5 in this example includes;

[0132] Step S51: Send commands according to priority via CAN bus, calculate the coordination parameters of the actuator when multiple commands are executed in coordination, monitor the execution status and progress of the action in real time, and trigger a reset and alarm when a jam or deviation occurs.

[0133] In this embodiment of the invention, correction control commands are sequentially issued via the CAN bus according to command priority. When multiple commands are executed in coordination, the coordination parameters of the actuator are calculated in advance to avoid action conflicts. The action status and progress of the actuator are monitored in real time. Sensor feedback data is used to determine whether there are abnormal situations such as jamming or exceeding tolerance. Once an abnormality occurs, the actuator is immediately reset and an alarm signal is issued to prevent the abnormality from expanding and affecting the excavation quality.

[0134] Step S52: After the correction command is executed, delay for 50ms, collect feedback data at a frequency of 200Hz for 1s, and after outlier removal, data alignment and coordinate mapping simplification preprocessing, take the average value to obtain the feedback data matrix and calculate the adjustment accuracy;

[0135] In this embodiment of the invention, a 50ms delay is made after the correction control command is executed to ensure the stability of the actuator. Feedback data is continuously collected for 1 second at a sampling frequency of 200Hz, for a total of 200 frames. The collected feedback data is simplified and preprocessed, including outlier removal, data alignment, and coordinate mapping. Then, the average value is taken to generate a feedback data matrix. Based on the comparison results between the feedback data matrix and the benchmark parameters, the adjustment accuracy of this correction is calculated to provide a basis for subsequent model fine-tuning.

[0136] Step S53: Trigger incremental fine-tuning based on the results of the precision grading, only adjust the weights of the fully connected layer and LSTM layer of the model, select the feedback data of the most recent 1000 frames as the fine-tuning dataset, and verify it through the test dataset after fine-tuning.

[0137] In this embodiment of the invention, the triggering conditions for incremental model fine-tuning include: adjustment accuracy < 70%, a cumulative total of 5 second-level precision samples, or changes in the working environment, such as soil type switching or weather changes; during fine-tuning, only the weights of the fully connected layers and LSTM layers of the model are adjusted, and the weight adjustment range is constrained as follows: fully connected layers ≤ ±5%, LSTM layers ≤ ±3%; the fine-tuning learning rate is set to 0.0001, and the fine-tuning rounds are 10 rounds, ensuring that the fine-tuning time is ≤ 5 minutes to avoid affecting the normal work progress; after the fine-tuning is completed, the model performance is verified through a test dataset to ensure that the adjustment is effective.

[0138] Step S54: Locate the source of the anomaly through a three-level anomaly classification mechanism, and execute prompts, audible and visual alarms, and emergency switching to manual control mode according to the anomaly level. After the anomaly is resolved, automatically switch back to automatic control mode.

[0139] In this embodiment of the invention, a three-level anomaly classification mechanism is adopted: Level 1 anomalies include single-frame data loss and adjustment accuracy of Level 2; Level 2 anomalies include sensor failure and adjustment accuracy of Level 3; Level 3 anomalies include actuator failure and model inference error. The classification process is as follows: Level 1 anomalies are automatically compensated by a data completion algorithm without affecting the operation; Level 2 anomalies trigger an audible and visual alarm and initiate emergency adjustment measures; Level 3 anomalies immediately trigger an emergency alarm and switch to manual control mode, locking the actuator to prevent safety accidents; after the anomaly is resolved, the system automatically switches back to automatic control mode.

[0140] Step S33 includes the following steps:

[0141] Step S331: Benchmark parameter subdivision and adaptation; subdivide the benchmark parameters according to every 0.5m excavation depth and the initial / middle / final stages of excavation, and preset corresponding parameters for square hole side lengths of 0.5-1.5m and depths of 2-5m respectively;

[0142] Step S332: Dynamic adjustment of reference parameters; Based on the actual contour of the foundation pit identified by the visual sensor, when the excavation deviation of a certain side wall is >10cm, the reference posture parameters of the corresponding side excavation are automatically adjusted, with an adjustment range ≤0.5°.

[0143] In this embodiment of the invention, the core of setting the benchmark parameters is to achieve dual precise matching of "subdivided adaptation + dynamic adaptation". The excavation depth is divided into intervals of 0.5m. Combined with the different operational requirements of the initial outline forming stage, the middle depth advancement stage, and the final accuracy correction stage, corresponding benchmark posture parameters are preset for the conventional engineering range of square hole side length 0.5-1.5m and depth 2-5m. This ensures that excavation operations of different specifications and progress have a targeted benchmark basis. Based on the actual outline of the foundation pit identified in real time by the visual sensor, a deviation monitoring mechanism is established. When the excavation deviation of a certain side wall exceeds 10cm, the system automatically locates the deviation side and adjusts the excavation benchmark posture parameters of the corresponding side. The adjustment range is strictly controlled to be ≤0.5°. This avoids the excavation trajectory from changing abruptly due to excessive benchmark adjustment range and can correct the outline deviation in time, ensuring the core accuracy indicators such as the verticality and side length consistency of the square hole foundation pit.

[0144] Step S53 includes the following steps:

[0145] Step S531: Fine-tuning trigger conditions; model fine-tuning is triggered when the adjustment accuracy is <70%, when there are a total of 5 second-level accuracy samples, or when the working environment changes.

[0146] Step S532: Fine-tune parameter constraints; the adjustment range of the weights of the fully connected layer is ≤ ±5%, the adjustment range of the weights of the LSTM layer is ≤ ±3%, the fine-tuning learning rate is set to 0.0001, the fine-tuning rounds are 10, and the fine-tuning time is ≤ 5 min.

[0147] In this embodiment of the invention, incremental model fine-tuning aims to achieve a balance between "precise adaptation and uninterrupted operation." When the adjustment accuracy is <70%, it indicates that the model's current adaptability is insufficient and needs timely optimization. When a total of 5 samples with secondary accuracy are accumulated, it indicates a trend of deviation accumulation, requiring early intervention. When the operating environment changes, it is necessary to quickly adapt to the new scenario: secondly, only the weights of the fully connected layer and LSTM layer are adjusted to avoid inference delays caused by overall model reconstruction; the adjustment range of the fully connected layer weights is ≤±5%, and the LSTM layer weights are ≤±3% to prevent model instability caused by sudden weight changes; the fine-tuning learning rate is set to 0.0001 and the number of rounds is 10 to ensure smooth convergence of the fine-tuning process, and the fine-tuning time is limited to ≤5 minutes to avoid affecting the normal excavation operation progress; at the same time, the feedback data of the most recent 1000 frames is used as the fine-tuning dataset to ensure that the model optimization is based on the latest actual operating conditions and improve real-time adaptation capabilities.

[0148] Step S54 includes the following steps:

[0149] Step S541: Anomaly classification criteria; Level 1 anomaly is single frame data loss, and Level 2 adjustment accuracy is also anomaly; Level 2 anomaly is sensor failure, and Level 3 adjustment accuracy is also anomaly; Level 3 anomaly is actuator failure, and Model inference error.

[0150] Step S542: Graded processing flow; Level 1 anomalies are automatically compensated, Level 2 anomalies trigger audible and visual alarms and emergency measures, Level 3 anomalies trigger emergency alarms and switch to manual control mode, while locking the actuator.

[0151] In this embodiment of the invention, anomaly handling follows the principle of "tiered response + precise prevention and control," balancing operational safety and continuity. Anomalies are categorized into three levels based on their impact and risk level: Level 1 anomalies involve single-frame data loss or Level 2 accuracy adjustments, with a small impact range and quick recovery; Level 2 anomalies involve sensor malfunctions or Level 3 accuracy adjustments, affecting data acquisition or correction effectiveness and requiring vigilance; Level 3 anomalies involve actuator malfunctions or model inference errors, directly threatening operational safety or causing excavation failure, posing the highest risk. Specifically, the core of this approach lies in establishing differentiated handling procedures: Level 1 anomalies are automatically compensated for through data completion algorithms, requiring no manual intervention and ensuring uninterrupted operation; Level 2 anomalies trigger audible and visual alarms and simultaneously initiate emergency adjustment measures, such as switching to backup sensor data or temporarily adjusting correction strategies, reminding operators to promptly investigate; Level 3 anomalies immediately trigger an emergency alarm, simultaneously switching to manual control mode and locking the actuator to prevent malfunctions that could cause equipment damage or safety accidents. After the anomaly is investigated and repaired, the system automatically switches back to automatic control mode, ensuring operational safety and process continuity. Furthermore, fault tree analysis is used throughout the process to locate the anomaly source, improving fault handling efficiency.

[0152] This invention also provides a deep learning-based real-time posture correction and control system for excavating square hole foundation pits on excavators, used to execute the deep learning-based real-time posture correction and control method for excavating square hole foundation pits on excavators as described above. The system includes:

[0153] The multi-source data acquisition module is used to synchronously acquire bucket attitude parameters, pit contour data and relative position data through redundantly deployed attitude sensors, vision sensors and laser rangefinders, using a synchronous triggering mechanism and dual-link redundant transmission, forming a raw data stream with a check code, and supporting adaptive adjustment of acquisition frequency and hierarchical storage.

[0154] The data preprocessing and standardization module is used to perform outlier removal, adaptive Kalman filter noise filtering, and spatiotemporal alignment on the raw data stream, construct a three-dimensional rectangular coordinate system for the square hole foundation pit, and uniformly map and standardize the data into a standardized data matrix in the [0,1] interval;

[0155] The pose recognition and deviation calculation module is used to load the CNN-LSTM hybrid model trained and optimized by transfer learning, output the current pose parameters through multi-source data fusion, call the dynamically adapted baseline pose parameter library, calculate multi-dimensional pose deviation vectors and optimize filtering.

[0156] The correction control command generation module is used to obtain targeted correction control commands based on deviation priority sorting through a two-dimensional fuzzy PID controller. After verification and format conversion by the excavator kinematic model, the commands are sent to the electronic control execution unit according to priority.

[0157] The execution feedback optimization module is used to monitor the command execution status in real time, collect and preprocess feedback data, calculate and adjust accuracy and trigger incremental fine-tuning of the model. It adopts a three-level anomaly classification and processing mechanism to realize fault location, alarm and mode switching, and build a closed-loop correction control system.

[0158] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for real-time correction and control of the excavation posture of an excavator in a square hole foundation pit for utility poles, based on deep learning, characterized in that... Includes the following steps: Step S1: Acquire real-time multi-source data; Using the attitude sensor, vision sensor and laser rangefinder integrated into the excavator's electronic control system, a unified synchronous triggering mechanism is used to synchronously collect the excavator bucket attitude parameters, pit outline data and relative position data of the bucket and square hole reference trajectory during the excavation process; The raw data stream is formed based on CAN bus + Ethernet dual-link redundant transmission; Step S2: Data preprocessing and standardization; outlier removal, noise filtering and data alignment are performed on the raw data stream, and the attitude parameters and position data are uniformly mapped to the preset three-dimensional rectangular coordinate system of the square hole foundation pit to obtain a standardized data matrix. Step S3: Deep learning model-driven attitude recognition and deviation calculation; Input the standardized data matrix into the CNN-LSTM hybrid model trained by transfer learning and quantized and optimized, output the current excavation attitude parameters of the bucket through multi-source data fusion, and compare it with the dynamically adapted square hole foundation pit excavation benchmark attitude parameters to obtain multi-dimensional attitude deviation values. Step S4: Generation of correction control commands; Based on the priority ranking of attitude deviation values, targeted attitude correction control commands are generated using a two-dimensional fuzzy PID control algorithm; Step S5: Real-time execution and feedback optimization; The correction control command is sent to the excavator's electronic control execution unit, the command execution status is monitored in real time, the corrected posture data is collected and simplified preprocessed to generate a feedback data matrix, the adjustment accuracy triggers the model for incremental fine-tuning, and an anomaly classification handling mechanism is adopted to ensure operational safety.

2. The method for real-time correction and control of excavator pole square hole foundation pit excavation posture based on deep learning according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Obtain the excavator bucket attitude parameters, pit outline data, and relative position data between the bucket and the square hole reference trajectory; Step S12: Use CAN bus to transmit attitude and laser data, and Ethernet to transmit vision data. Each link has a dedicated channel. The multi-sensor protocol is uniformly converted to CANopen protocol through the built-in protocol conversion module of the PLC controller. Step S13: Generate a unified synchronous trigger signal through the PLC controller, and the acquisition frequency is adaptively adjusted according to the excavation depth and operation status.

3. The method for real-time correction and control of excavator pole square hole foundation pit excavation posture based on deep learning according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Identify outliers and obtain outlier data; Step S22: Perform targeted noise filtering on attitude parameters, laser distance data, and visual image data; Step S23: Perform time alignment on the multi-source data to obtain aligned data; Step S24: Construct a three-dimensional rectangular coordinate system with the geometric center of the square hole foundation pit as the origin, and map the aligned data to this coordinate system to obtain a standardized data matrix.

4. The method for real-time correction and control of excavator pole square hole foundation pit excavation posture based on deep learning according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Construct a million-level dataset covering different soil types, excavation depths and working environments, perform data augmentation, train a CNN-LSTM hybrid model using transfer learning strategy, load pre-trained weights through ResNet50, and perform quantization processing through TensorRTINT8; Step S32: Integrate multi-sensor data to obtain attitude parameters, cross-validate them with laser and vision data, and determine the parameters to be valid if the verification error is ≤1mm; Step S33: Based on the design dimensions of the square hole foundation pit, soil type, and detailed benchmark posture parameters of the excavation stage, construct a local SQLite parameter library, and dynamically adjust the benchmark parameters based on the actual excavation contour during operation; Step S34: Calculate the angle deviation, position deviation, depth deviation and comprehensive deviation respectively; filter the instantaneous deviation and integrate them into a three-dimensional deviation vector output with confidence label.

5. The method for real-time correction and control of excavator pole square hole foundation pit excavation posture based on deep learning according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Construct a two-dimensional fuzzy controller with the inputs being the deviation value e and the deviation change rate ec, and the outputs being the PID parameter corrections ΔKp, ΔKi, and ΔKd. Formulate a 7×7 fuzzy control rule table. Step S42: Obtain the bucket slewing angle correction command based on the angle deviation, obtain the stick extension length correction command based on the position deviation, and obtain the boom lifting height correction command based on the depth deviation and pitch angle deviation; Step S43: Instruction optimization and verification; sort instructions according to priority order of angle deviation > depth deviation > position deviation, verify the validity of instructions, convert them into a format recognizable by the electronic control unit and add a check code, with instruction transmission delay ≤ 5ms.

6. The method for real-time correction and control of excavator pole square hole foundation pit excavation posture based on deep learning according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Send commands according to priority via CAN bus, calculate the coordination parameters of the actuator when multiple commands are executed in coordination, monitor the execution status and progress of the action in real time, and trigger a reset and alarm when a jam or deviation occurs. Step S52: After the correction command is executed, delay for 50ms, collect feedback data at a frequency of 200Hz for 1s, and after outlier removal, data alignment and coordinate mapping simplification preprocessing, take the average value to obtain the feedback data matrix and calculate the adjustment accuracy; Step S53: Trigger incremental fine-tuning based on the results of the precision grading, only adjust the weights of the fully connected layer and LSTM layer of the model, select the feedback data of the most recent 1000 frames as the fine-tuning dataset, and verify it through the test dataset after fine-tuning. Step S54: Locate the source of the anomaly through a three-level anomaly classification mechanism, and execute prompts, audible and visual alarms, and emergency switching to manual control mode according to the anomaly level. After the anomaly is resolved, automatically switch back to automatic control mode.

7. The method for real-time correction and control of excavator pole square hole foundation pit excavation posture based on deep learning according to claim 4, characterized in that, Step S33 includes the following steps: Step S331: Benchmark parameter subdivision and adaptation; subdivide the benchmark parameters according to every 0.5m excavation depth and the initial / middle / final stages of excavation, and preset corresponding parameters for square hole side lengths of 0.5-1.5m and depths of 2-5m respectively; Step S332: Dynamic adjustment of reference parameters; Based on the actual contour of the foundation pit identified by the visual sensor, when the excavation deviation of a certain side wall is >10cm, the reference posture parameters of the corresponding side excavation are automatically adjusted, with an adjustment range ≤0.5°.

8. The method for real-time correction and control of the excavation posture of an excavator for a square hole foundation pit based on deep learning, as described in claim 6, is characterized in that... Step S53 includes the following steps: Step S531: Fine-tuning trigger conditions; model fine-tuning is triggered when the adjustment accuracy is <70%, when there are a total of 5 second-level accuracy samples, or when the working environment changes. Step S532: Fine-tune parameter constraints; the adjustment range of the weights of the fully connected layer is ≤ ±5%, the adjustment range of the weights of the LSTM layer is ≤ ±3%, the fine-tuning learning rate is set to 0.0001, the fine-tuning rounds are 10, and the fine-tuning time is ≤ 5 min.

9. The method for real-time correction and control of the excavator pole square hole foundation pit excavation posture based on deep learning according to claim 6, characterized in that, Step S54 includes the following steps: Step S541: Anomaly classification criteria; Level 1 anomaly is single frame data loss, and Level 2 adjustment accuracy is also anomaly; Level 2 anomaly is sensor failure, and Level 3 adjustment accuracy is also anomaly; Level 3 anomaly is actuator failure, and Model inference error. Step S542: Graded processing flow; Level 1 anomalies are automatically compensated, Level 2 anomalies trigger audible and visual alarms and emergency measures, Level 3 anomalies trigger emergency alarms and switch to manual control mode, while locking the actuator.

10. A real-time posture correction control system for excavating square hole foundation pits using excavators based on deep learning, characterized in that, For executing the deep learning-based real-time correction and control method for excavator pole square hole foundation pit excavation posture as described in any one of claims 1-9, the system comprises: The multi-source data acquisition module is used to synchronously acquire bucket attitude parameters, pit contour data and relative position data through redundantly deployed attitude sensors, vision sensors and laser rangefinders, using a synchronous triggering mechanism and dual-link redundant transmission, forming a raw data stream with a check code, and supporting adaptive adjustment of acquisition frequency and hierarchical storage. The data preprocessing and standardization module is used to perform outlier removal, adaptive Kalman filter noise filtering, and spatiotemporal alignment on the raw data stream, construct a three-dimensional rectangular coordinate system for the square hole foundation pit, and uniformly map and standardize the data into a standardized data matrix in the [0,1] interval; The pose recognition and deviation calculation module is used to load the CNN-LSTM hybrid model trained and optimized by transfer learning, output the current pose parameters through multi-source data fusion, call the dynamically adapted baseline pose parameter library, calculate multi-dimensional pose deviation vectors and optimize filtering. The correction control command generation module is used to obtain targeted correction control commands based on deviation priority sorting through a two-dimensional fuzzy PID controller. After verification and format conversion by the excavator kinematic model, the commands are sent to the electronic control execution unit according to priority. The execution feedback optimization module is used to monitor the command execution status in real time, collect and preprocess feedback data, calculate and adjust accuracy and trigger incremental fine-tuning of the model. It adopts a three-level anomaly classification and processing mechanism to realize fault location, alarm and mode switching, and build a closed-loop correction control system.

Images