A torque signal-based dewatering robot obstacle recognition and tool switching system

Abstract

The application discloses a kind of based on torque signal's dredging robot obstacle identification and tool switching system, it is related to artificial intelligence technical field, including: mechanical arm joint torque acquisition module, for real-time acquisition torque signal;Torque signal pre-processing module is for torque signal is segmented to sliding window, removes direct current component, filter denoising and normalization processing;Obstacle mechanics feature extraction module is for extracting time domain, frequency domain and time-frequency domain multidimensional mechanics feature;Obstacle type identification module is for inputting mechanics feature vector into pre-trained classification model to output obstacle type;Tool switching decision module is for determining optimal dredging tool according to obstacle type and generating switching instruction.The application uses torque signal to replace visual information, can independently identify obstacle type in turbid water body and switch dredging tool, realizes the intelligent dredging operation of unmanned intervention.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a system for obstacle identification and tool switching in a dredging robot based on torque signals. Background Technology

[0002] A core challenge in the actual operation of drainage pipe dredging robots is that operators can remotely observe the situation inside the pipes via cameras, but the water inside the drainage pipes is turbid and visibility is extremely low. Traditional visual sensors become almost completely ineffective after the sludge is stirred up during dredging operations. Operators cannot accurately determine whether the material in front of the cutter head is loose sludge, hardened lumps, tree root fibers, or steel bars and gravel. By the time the motor overload alarm is detected, the cutter head has already been entangled by tree roots or the blades are severely worn.

[0003] This invention aims to provide a method for obstacle identification and autonomous tool switching control based on robotic arm torque signals, in order to solve the problems in the prior art where dredging robots rely on visual perception, cannot autonomously determine obstacle types, and rely on manual remote control for tool switching. This method enables dredging robots to autonomously identify obstacle types, actively escape from obstacles, and autonomously switch dredging tools in turbid water. Summary of the Invention

[0004] The main objective of this invention is to disclose an obstacle recognition and tool switching system for a dredging robot based on torque signals, comprising: Robotic arm joint torque acquisition module: includes torque sensors installed at each joint of the robotic arm for real-time acquisition of torque signals; The robotic arm joint torque acquisition module includes: a torque sensor, a torque signal acquisition unit, and a shaft angle encoder unit; the torque sensor is integrated into the harmonic reducer or the rear end of the motor of the robotic arm joint to directly measure the actual torque driving the joint rotation; the torque signal acquisition unit collects the torque values ​​of each joint of the robotic arm at a preset sampling frequency; and the shaft angle encoder collects the rotation angle of the cutter spindle in real time.

[0005] Torque signal preprocessing module: used to perform sliding window segmentation, DC component removal, filtering and noise reduction, and normalization on the acquired torque signal to generate a preprocessed torque signal; The torque signal preprocessing module includes: Sliding window segmentation unit: Overlapping sliding windows are used to process real-time data streams; DC component removal function unit: Subtract the mean value from the torque sequence of each joint within the window to eliminate the effects of tool weight and steady-state bias; Low-pass filter function unit: Uses a low-pass filter to filter the torque signal after DC removal; Normalization unit: Used to normalize the subsequently extracted features to eliminate the influence of different units on the classification results.

[0006] Obstacle mechanical feature extraction module: used to extract the mechanical feature vector of the obstacle; the mechanical feature vector is obtained by extracting multi-dimensional features in the time domain, frequency domain and time-frequency domain based on the preprocessed torque signal. Specifically, the obstacle mechanical feature extraction module normalizes the extracted mechanical features to construct a multidimensional mechanical feature vector. Among them, the mechanical characteristics include: mean torque, torque variance, torque kurtosis, autocorrelation periodicity, zero-crossing rate, phase synchronization characteristics, and entanglement index.

[0007] Dynamic entanglement evolution determination module: used to identify continuous entanglement of tree root fibers, tool stall, periodic pulling, idling slippage and continuous load accumulation.

[0008] Tool Health Status Assessment Module: Used to analyze the energy in the high-frequency region of the spectrum, determine the health status of the tool, and assess the wear status of the dredging tool.

[0009] Obstacle type identification module: This module takes the mechanical feature vectors as input to a pre-trained classification model and outputs the obstacle type. Specifically, the obstacle type identification module adopts a one-to-one multi-classification strategy to train a binary support vector machine sub-classifier. The sub-classifier uses a Gaussian radial basis function as the kernel function to distinguish different obstacle categories. The obstacle type identification module includes: an obstacle category definition unit and a classifier unit; Specifically, the obstacle category definition is implemented based on the actual situation of urban underground drainage pipe network dredging operations; the classifier unit is a mapping model that uses support vector machine to realize the mapping between multi-dimensional mechanical feature vectors and obstacle categories; Furthermore, the categories of obstacles include: soft deposits, hardened lumps, invading plant roots, and gravel.

[0010] Tool switching decision module: Generates tool switching instructions and active escape action instructions based on obstacle type and entanglement status; Specifically, the tool switching decision module receives obstacle category information, establishes a mapping table between obstacle types and dredging tools, and executes dredging tool and escape action commands.

[0011] Command sending and execution interface module: used to convert tool switching commands into control signals; the command sending and execution interface module includes command format definition, communication protocol and reliability guarantee, and physical interface unit of execution hardware; Specifically, the communication protocol and reliability guarantee are achieved through a communication confirmation and timeout retransmission mechanism. The communication confirmation and timeout retransmission mechanism means that after the instruction is sent, the system waits for the receiver's response frame. If no response frame is received within a preset timeout period, the instruction is automatically retransmitted and retried a preset number of times. After three failures, a communication fault alarm is triggered and the system enters a safe standby state.

[0012] According to the present invention, contact force information is collected by torque sensors at the joints of the robotic arm, and the type of obstacle in front is identified by analyzing the time-domain characteristics of the torque signal. Obstacle recognition can be achieved in turbid water without relying on visual information. This overcomes the limitation of traditional visual cameras in being unable to effectively perceive the environment in extremely low visibility conditions. It realizes an unmanned operation mode in which the robot can autonomously move along the pipeline and perform dredging tasks after being remotely started by the operator, thereby improving the automation level and efficiency of dredging operations. Attached Figure Description

[0013] Figure 1 This is a system architecture diagram according to an embodiment of the present invention. Detailed Implementation

[0014] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0015] like Figure 1 As shown, the main objective of this invention is to disclose an obstacle identification and tool switching system for a dredging robot based on torque signals, comprising: P100 robotic arm joint torque acquisition module: Torque sensors are installed at each joint of the robotic arm to collect the real torque signals that drive the joints to rotate in real time; The P100 robotic arm joint torque acquisition module includes: a P110 torque signal acquisition unit, a P120 torque sensor unit, and a P130 shaft angle encoder unit. Specifically, the P110 torque signal acquisition unit uses a preset sampling frequency. Torque values ​​of each joint of the robotic arm Collect data; in, For the i-th joint in the i-th... The actual torque value at each sampling time point, where i (i=1, 2, 3…n) is the joint number. is the sampling sequence number, and n is the total number of robotic arm joints.

[0016] The P120 torque sensor is integrated into the harmonic reducer or the rear end of the motor of the robotic arm joint to directly measure the actual torque that drives the joint to rotate. The P130 shaft angle encoder is used to acquire the rotation angle of the auger spindle in real time. Calculate the corresponding rotation phase : ; in, The rotation phase corresponding to the k-th sampling time; The sampling sequence number; The sampling frequency; This is the rotation frequency of the auger spindle.

[0017] P200 Torque Signal Preprocessing Module: Used to perform sliding window segmentation, filtering, noise reduction, and detrending processing on the acquired torque signal to generate a preprocessed torque signal; The P200 torque signal preprocessing module includes: a P210 sliding window segmentation unit, a P220 DC component removal unit, a P230 low-pass filtering unit, and a P240 normalization processing unit.

[0018] The specific steps for the P210 sliding window segmentation unit are as follows: Step 1: Use overlapping sliding windows to process real-time data streams; Step 2: Define window length (The time step can be changed in practical applications), sliding step size For the j-th window, the set of sample indices it contains is: ; in, Let N be the j-th sliding window, S be the window length, S be the sliding step size, j be the window number, and k be the sampling sequence number. Specifically, the window length N is preferably used to cover 1 to 3 seconds of real-time data to ensure that the complete mechanical behavior during the obstacle contact process can be preserved; the sliding step size S is preferably set to the number of sampling points corresponding to 0.5 seconds to improve the real-time response capability of the system.

[0019] P220 DC component removal unit: For the torque sequence of each joint within the window, subtract its mean value to eliminate the effects of tool weight and steady-state bias; the specific formula is: ; in, Torque value after DC removal Let be the sampled value of the i-th joint at the k-th sampling time. Let N be the original torque value of the i-th joint at the m-th sampling time, N be the window length, and k and m be the sampling sequence numbers.

[0020] P230 Low-Pass Filter Unit: A low-pass filter is used to filter the torque signal after DC removal. The filter is discretized using the bilinear transform method. The digital filter difference equation is as follows: ; in, For input torque signal, This is the filtered torque signal. For the filter forward coefficients (m=0,1,2,3,4). For the filter feedback coefficients (m=1,2,3,4); Specifically, and Typical values ​​are: =[0.0934,0.3759,0.5638,0.3759,0.0934], =[1, -0.4860, 0.5262, -0.1178, 0.0276].

[0021] P240 Normalization unit: Used to normalize the subsequently extracted features to eliminate the influence of different units on the classification results; For any feature parameter Its normalization formula is: ; in, The normalized value. The original value, The maximum value of this feature in the training set. This is the minimum value of this feature in the training set.

[0022] P300 Obstacle Mechanical Feature Extraction Module: Used to extract the mechanical feature vectors of obstacles; The mechanical feature vector is obtained by extracting multi-dimensional features in the time domain, frequency domain, and time-frequency domain based on the preprocessed torque signal; Specifically, when the dredging robot comes into contact with different obstacles, the torque signal will exhibit significantly different time-domain statistical characteristics and frequency-domain distribution characteristics due to the differences in the material properties and geometric shapes of the obstacles; Among them, the mechanical characteristic quantities include: mean torque. Torque variance Torque kurtosis K, autocorrelation periodicity Zero-crossing rate (ZCR), phase synchronization characteristics, and entanglement index (WI): 1) Mean torque This reflects the average load level borne by the robotic arm within the window and is a fundamental indicator for measuring the overall resistance of an obstacle. The formula is: ; in, The average torque. The torque value at the k-th sample point. This is the window length.

[0023] 2) Torque variance : Measures the degree of fluctuation of the torque signal around its mean, reflecting the smoothness of the contact process. The formula is: .

[0024] 3) Torque kurtosis : A fourth-order statistic describing the distribution pattern of a signal, highly sensitive to the presence of impulsive spikes in the signal. The formula is: ; in, It is a very small positive number (e.g., take 10). -6 ), used to prevent division by zero errors when the variance is close to zero; when season =3.

[0025] 4) Autocorrelation periodicity When fibrous materials such as tree roots are wrapped around a auger, the wrapped fibers are stretched once for each rotation of the auger, resulting in a periodic fluctuation in the torque signal that is strictly corresponding to the rotation speed of the auger; the autocorrelation function can effectively extract the hidden periodic components in the signal. The specific steps are as follows: Step 1: Calculate the unbiased autocorrelation function of the torque signal. : ; in, Lagging The autocorrelation value of the step, =0, 1, 2, ..., Lmax, where Lmax is the maximum calculation lag step, covering at least one cutter rotation cycle; k is the sampling sequence number; and These are the torque values ​​corresponding to the k-th and (k+l)-th sampling points, respectively; Step 2: Normalize the autocorrelation function to obtain the normalized autocorrelation function. : ; in The normalized autocorrelation value is dimensionless and ranges from [-1, 1]. The autocorrelation value is zero. Step 3: In the normalized autocorrelation function sequence, find the lag number that satisfies the following conditions. ( >0): and ; ; If a match is found Then the autocorrelation periodic characteristics are: , The sampling frequency.

[0026] 5) Zero-crossing rate (ZCR): Measures the frequency at which a signal crosses above and below its mean, reflecting the proportion of high-frequency components in signal fluctuations. ; in, This is an indicator function; it takes the value 1 when the condition within the parentheses is true, and 0 otherwise; k is the sampling number. 6) Phase synchronization feature: When fibrous obstacles such as tree roots are wrapped around the surface of the auger, the auger will generate periodic pulling at a fixed angle position every time it rotates, resulting in the peak torque being synchronized with the rotation phase of the auger. Specifically, the phase corresponding to the peak torque in the nth rotational cycle is defined as: Define the phase shift between two consecutive cycles. for: ; Where n is the rotation period number; Furthermore, when multiple consecutive periods satisfy If so, it is determined that a stable fiber entanglement state exists. This is the phase drift threshold.

[0027] 7) Entanglement Index (WI): This comprehensively describes the degree of obstacle entanglement, constructs an entanglement index, and determines the current fiber entanglement state of the obstacle. The specific steps are as follows: Step 1: Construct the entanglement index, the specific formula is as follows: ; Where WI is the entanglement index. The peak amplitude of the period. The periodic correlation coefficient; This is the average phase drift. , , These are the weighting coefficients; Step 2: Determine the current fiber entanglement state of the obstacle. The determination formula is: when... The system determines that the current obstacle is in a state of fiber entanglement. The threshold for determining entanglement.

[0028] P400 Dynamic Wrapping Evolution Determination Module: Used to identify continuous entanglement of tree root fibers, tool stall, periodic pulling, idling slippage, and continuous load accumulation; Specifically, when the obstacle has a fibrous structure, the torque fluctuation manifests as an instantaneous high load, which gradually forms a stable periodic phase-locking phenomenon as the degree of entanglement increases. Furthermore, the trend of the entanglement index WI changing over time is monitored in real time. ,when If the entanglement continues for more than a preset time, it is determined that the current entanglement state is deteriorating; in, This represents the rate of change of the entanglement index.

[0029] P500 Tool Health Status Assessment Module: Used to analyze the energy in the high-frequency region of the spectrum, determine the health status of the tool, and assess the wear status of the dredging tool; Specifically, after long-term wear, the high-frequency energy generated by the cutting impact of a cutting tool will significantly decrease. The tool's health status can be determined by analyzing the energy in the high-frequency region of the spectrum. The formula is: ; when When determining that the tool is in a worn or dull state, The wear determination threshold; in, High-frequency energy, The torque signal spectrum; This is the high-frequency cutoff frequency; This represents the maximum analysis frequency.

[0030] P600 obstacle type identification module: used to input mechanical feature vectors into a pre-trained classification model and output obstacle type; The obstacle type identification module includes: P610 obstacle category definition unit and P620 classifier unit; The P610 obstacle category definition is implemented based on the actual situation of urban underground drainage pipe network dredging operations, and obstacles are divided into four categories, as shown in Table 1: Table 1

[0031] The P620 classifier unit is a mapping model that uses support vector machines to map feature vectors to obstacle categories. Specifically: 1) Multi-classification strategy: A one-versus-one multi-classification strategy is adopted, and the system is trained on four types of obstacles. Each binary SVM subclassifier is responsible for distinguishing between two specific categories (e.g., category 1 and category 2, category 1 and category 3).

[0032] 2) Kernel Function: Each binary SVM uses a Gaussian radial basis function (RBF) as its kernel function, which has the following form: ; in, The kernel function value measures the similarity between two feature vectors. The kernel function scaling parameter, It is a Euclidean distance.

[0033] 3) Training and deployment: The training phase is completed offline before the actual deployment of the robot. After training, a classification model is formed and deployed to the robot controller. Specifically, labeled datasets ,in Let be the feature vector of the m-th sample. For category labels, M is the total number of samples, and m is the sample number.

[0034] 4) The specific steps of the online classification process are as follows: Step 1: During actual robot operation, every 0.5 seconds (sliding step) =0.5 s) generates a new feature vector F, which is simultaneously fed into 6 pre-trained binary SVM sub-classifiers; for the sub-classifier responsible for distinguishing between classes p and q, the decision function value is for: ; Step 2: This is the set of support vectors for the sub-classifier. For Lagrange multipliers, The support vectors are labeled with their class (class p is labeled as +1 and class q is labeled as -1 during training). The feature vector to be classified The feature vectors corresponding to the support vectors. Why kernel functions? For classification bias items; Step 3: Give a vote based on the sign of the decision function: ; Step 4: After the 6 sub-classifiers vote independently, count the number of votes received for each category and output the obstacle category. The category that received the most votes.

[0035] P700 Tool Switching Decision Module: Generates tool switching commands and active escape action commands based on obstacle type and entanglement status; includes: P710 Tool Switching Unit and P720 Active Escape Control Unit.

[0036] P710 Tool Switching Unit: By receiving obstacle category information, it establishes a mapping table between obstacle types and dredging tools, as shown in Table 2. Table 2

[0037] Specifically, the mapping table is embedded in the robot's decision-making process. When the classifier outputs a category... In this case, the robot can immediately look up a table to determine "which tool should be used," and the specific formula is as follows: ; Where M is the toolset.

[0038] P720 Active Escape Control Unit: Used to determine the current fiber entanglement status of the obstacle and execute active escape actions; Specifically, when detected At that time, the robot performs an active escape maneuver. The specific status is as follows: 1) Reverse unwinding: Control the auger to perform a short-term reverse rotation; 2) Pulse oscillation desorption: Applying a periodic torsional torque to the articulated arm; 3) Axial retraction: The robotic arm moves backward along the axis, retracting a preset distance d.

[0039] P800 Instruction Sending and Execution Interface Module: Used to convert tool switching instructions into control signals, including: P810 instruction format definition, P820 communication protocol and reliability guarantee, and P830 physical interface unit for executing hardware.

[0040] The P810 instruction format definition refers to the tool switching instructions that use the structured field data frame format definition, as shown in Table 3: Table 3

[0041] P820 ​​Communication Protocol and Reliability: Ensures that commands are sent using the industrial fieldbus protocol. After each command is sent, the sender waits for the receiver's acknowledgment frame (ACK), which contains the reception status code. 0x01: The instruction was successfully received and executed; 0x02: Verification error, request resend; 0x03: The executing agency is busy, and instructions are queued; 0x04: Instruction is illegal and cannot be executed; Specifically, if the sender does not receive an ACK within a preset timeout period (e.g., 100 ms), it will automatically resend the command, with a maximum of 3 retries; after 3 failures, a communication failure alarm will be triggered, and the robot will enter a safe standby state.

[0042] The P830 executes the physical interface of the hardware: it sends tool switching commands to the controller via the CAN bus, and the controller drives the solenoid valve or motor to complete the locking / releasing action of the tool head.

[0043] Ultimately, through the P100 robotic arm joint torque acquisition module, P200 torque signal preprocessing module, P300 obstacle mechanical feature extraction module, P400 dynamic entanglement state evolution determination module, P500 tool health status assessment module, P600 obstacle type identification module, P700 tool switching decision module, and P800 command sending and execution interface module, the dredging robot completes a complete closed loop from the acquisition of the original torque signal to the output of the final control command, realizing autonomous perception, autonomous identification, autonomous escape from obstacles, and dynamic tool switching in complex underground obstacle environments.

[0044] This invention, based on torque signal analysis, replaces the obstacle identification function of traditional visual sensors in turbid water, solving problems such as the failure of existing dredging robots relying on visual perception in turbid water, their inability to autonomously determine obstacle types, and the reliance on manual remote control for tool switching. It enables dredging robots to autonomously identify obstacle types and autonomously switch dredging tools in turbid water, significantly improving the intelligence level and unmanned operation capability of dredging robots.

[0045] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A system for obstacle recognition and tool switching in a dredging robot based on torque signals, characterized in that, include: Robotic arm joint torque acquisition module: includes torque sensors installed at each joint of the robotic arm for real-time acquisition of torque signals; Torque signal preprocessing module: used to perform sliding window segmentation, DC component removal, filtering and noise reduction, and normalization on the acquired torque signal to generate a preprocessed torque signal; Obstacle mechanical feature extraction module: used to extract the mechanical feature vector of the obstacle; the mechanical feature vector is obtained by extracting multi-dimensional features in the time domain, frequency domain and time-frequency domain based on the preprocessed torque signal. Dynamic entanglement state evolution determination module: used to identify continuous entanglement of tree root fibers, tool stall, periodic pulling, idling slippage and continuous load accumulation; Tool health status assessment module: used to analyze the energy in the high-frequency region of the spectrum, determine the health status of the tool, and assess the wear status of the dredging tool; Obstacle type identification module: This module takes the mechanical feature vectors as input to a pre-trained classification model and outputs the obstacle type. Tool switching decision module: Generates tool switching instructions and active escape action instructions based on obstacle type and entanglement status; Command sending and execution interface module: used to convert tool switching commands into control signals; The instruction sending and execution interface module includes: instruction format definition, communication protocol and reliability guarantee, and physical interface unit for execution hardware.

2. The obstacle identification and tool switching system for a dredging robot based on torque signals as described in claim 1, characterized in that, The robotic arm joint torque acquisition module includes: a torque sensor, a torque signal acquisition unit, and a shaft angle encoder; Among them, the torque sensor is integrated into the harmonic reducer or the rear end of the motor of the robotic arm joint to directly measure the real torque driving the joint rotation; the torque signal acquisition device collects the torque values ​​of each joint of the robotic arm at a preset sampling frequency; and the shaft angle encoder collects the rotation angle of the auger spindle in real time.

3. The obstacle identification and tool switching system for a dredging robot based on torque signals as described in claim 1, characterized in that, The torque signal preprocessing module includes: Sliding window segmentation unit: Overlapping sliding windows are used to process real-time data streams; DC component removal function unit: Subtract the mean value from the torque sequence of each joint within the window to eliminate the effects of tool weight and steady-state bias; Low-pass filter function unit: Uses a low-pass filter to filter the torque signal after DC removal; Normalization unit: Used to normalize the subsequently extracted features to eliminate the influence of different units on the classification results.

4. The obstacle identification and tool switching system for a dredging robot based on torque signals as described in claim 1, characterized in that, The obstacle mechanical feature extraction module normalizes the extracted mechanical features to construct a multidimensional mechanical feature vector.

5. The obstacle identification and tool switching system for a dredging robot based on torque signals as described in claim 4, characterized in that, The mechanical characteristics include: mean torque, torque variance, torque kurtosis, autocorrelation periodicity, zero-crossing rate, phase synchronization characteristics, and entanglement index.

6. The obstacle identification and tool switching system for a dredging robot based on torque signals as described in claim 1, characterized in that, The obstacle type identification module adopts a one-to-one multi-classification strategy to train a binary support vector machine sub-classifier. The sub-classifier uses a Gaussian radial basis function as the kernel function to distinguish different obstacle categories.

7. The obstacle identification and tool switching system for a dredging robot based on torque signals as described in claim 1, characterized in that, The obstacle type identification module includes: an obstacle category definition and a classifier unit; Among them, the obstacle category definition is implemented based on the actual situation of urban underground drainage pipe network dredging operations; the classifier unit is a mapping model that uses support vector machine to realize the mapping between multi-dimensional mechanical feature vectors and obstacle categories.

8. The obstacle identification and tool switching system for a dredging robot based on torque signals as described in claim 7, characterized in that, The categories of obstacles include: soft silt, hardened lumps, invading plant roots, and gravel.

9. The obstacle identification and tool switching system for a dredging robot based on torque signals as described in claim 1, characterized in that, The tool switching decision module receives obstacle category information, establishes a mapping table between obstacle types and dredging tools, and executes dredging tool and escape action commands.

10. The obstacle identification and tool switching system for a dredging robot based on torque signals as described in claim 1, characterized in that, The communication protocol and reliability guarantee are achieved through communication confirmation and timeout retransmission mechanisms. The communication confirmation and timeout retransmission mechanism refers to the following: after the instruction is sent, the system waits for the receiver's response frame. If no response frame is received within the preset timeout period, the instruction is automatically retransmitted and retried a preset number of times. After three failures, a communication fault alarm is triggered and the system enters a safe standby state.

Images