An industrial robot adaptive control system with fusion situation prediction mechanism

Abstract

This invention discloses an adaptive control system for industrial robots that integrates a situation prediction mechanism. Specifically, it relates to the field of situation prediction based on feedforward networks and fuzzy logic. During the execution of the industrial robot, state input data is continuously sampled and input into a feedforward neural network capable of mapping time-series features to obtain a predicted output of the robot's task trend. By constructing a dynamic entropy deviation coordination mechanism between the feedforward neural network and the fuzzy logic inference model, modal attribution conflict identification and control coordination for critical states of multimodal tasks are achieved. This ensures the continuity of the task response path and the stable operation of the control system within the fuzzy transition range, thereby solving the problem of modal misalignment and execution loss of control caused by the lack of entropy coupling between feedforward prediction output and fuzzy inference judgment in the control decision domain.

Description

Technical Field

[0001] This invention relates to the field of situation prediction technology based on feedforward networks and fuzzy logic, and more specifically, to an adaptive control system for industrial robots that integrates situation prediction mechanisms. Background Technology

[0002] In multimodal task execution scenarios, current industrial robot systems generally rely on feedforward networks to construct trend mappings of state sequences in order to predict the evolution direction of future tasks in advance and perform so-called situation prediction operations, thereby realizing the preloading of command responses and the forward scheduling of control rhythm. However, the confidence output of situation prediction is highly dependent on the prior distribution assumptions and input structure stability during training. In non-static or boundary-ambiguous working conditions, this prediction mechanism cannot provide a reliable judgment on the "uncertain region of state affiliation". At this time, the control system often introduces a fuzzy logic mechanism, which constructs a fuzzy membership function to make a fuzzy determination of the possibility of the robot's state affiliation among multiple control modes, so as to achieve a smooth transition between control strategies and a flexible connection of modal evolution paths.

[0003] However, when industrial robots face the critical transition range during multi-mode control switching, the following chain of problems often occurs:

[0004] The first-level problem is the structural conflict in modality attribution determination:

[0005] Current fuzzy logic systems often output fuzzy states with high membership overlap near the control mode transition boundary, meaning that a certain state simultaneously belongs to multiple modal rule systems (such as task A and task B). This state is essentially a "control fuzzy band". Due to the trend inference characteristics of the fixed structure, feedforward networks may output highly concentrated prediction results based on the short-term situation sequence, directly determining that the state should be assigned to a completely different modal task (such as task C), thus creating a structural discrepancy between the prediction results and the fuzzy determination.

[0006] The second problem is that the situation prediction output and the fuzzy inference output lack a unified decision domain:

[0007] Because the fuzzy logic control structure and the feedforward network prediction mechanism are independent of each other in terms of information flow structure and objective function design, the reasoning results of the two for the same state often cannot be effectively negotiated at the control management level. This results in the system generating multiple sets of mutually exclusive control commands for the same state. This separation of decision domains forces the control system to retain multiple control paths at the same time and carry out resource competitive scheduling.

[0008] The third layer of problems is the loss of control and responsibility, and the loss of focus in response:

[0009] The aforementioned multi-path divergence will ultimately lead to the blurring of the focus of control. Without establishing a unified consensus on entropy weight, the system will mistakenly regard uncertain states as high-confidence prediction results, resulting in unstable multi-directional scattering of control behavior. In severe cases, this may manifest as freezing of the execution system, failure of policy rollback, and disordered task response.

[0010] Based on the above, the existing problems with the technology can be summarized as follows:

[0011] During the evolution of multi-mode task situations, the industrial robot control system lacks an entropy weight coordination mechanism between the unidirectional predictive output of the feedforward network and the polymorphic membership reasoning of fuzzy logic, which leads to a structural misalignment in the state critical judgment, and in turn triggers the risk of systemic collapse due to the split of the control decision domain and the loss of control over the transfer of task response responsibility.

[0012] This problem reflects a fundamental gap in existing technologies regarding situation prediction and fuzzy logic-based collaborative control mechanisms. It is the most potentially destructive yet easily overlooked problem in current adaptive control systems. Summary of the Invention

[0013] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an adaptive control system for industrial robots that integrates a situation prediction mechanism. By constructing a dynamic entropy deviation coordination mechanism between a feedforward neural network and a fuzzy logic reasoning model, it achieves the identification of modal attribution conflicts and coordination of control rights for critical states of multimodal tasks. This ensures the continuity of the task response path and the stable operation of the control system within the fuzzy transition range, thereby solving the problems of modal misalignment and execution loss of control caused by the lack of entropy weight coupling between feedforward prediction output and fuzzy reasoning judgment in the control decision domain.

[0014] To achieve the above objectives, the present invention provides the following technical solution: an industrial robot adaptive control system integrating a situation prediction mechanism, comprising a situation prediction module, a classification module, a comparison module, a misalignment identification module, a buffer module, and a coordination module;

[0015] The situation prediction module is used to continuously sample and acquire state input data during the execution of an industrial robot, and input the state input data into a feedforward neural network that has the ability to construct a time-series feature mapping, so as to obtain the predicted output result of the robot task trend.

[0016] The attribution determination module is used to input the state input data from the situation prediction module into the fuzzy logic reasoning model based on the fuzzy membership function, and to perform fuzzy membership reasoning on the attribution probability of the current state of the industrial robot among multiple control modes, so as to obtain the fuzzy membership value of the current state under each control mode.

[0017] The comparison module is used to build a collaborative discrimination model. It compares the prediction output obtained from the situation prediction module with the fuzzy membership value obtained from the determination and attribution module to extract the difference information between the two in the modality determination results.

[0018] The misalignment identification module is used to calculate the control decision entropy deviation index value of the difference information, determine whether the current state of the industrial robot is in a modal critical misalignment state, and when the entropy deviation index value exceeds the preset threshold, it is marked as a conflict risk state.

[0019] In a preferred embodiment, the buffer module is used to activate the fuzzy priority control mechanism when a conflict risk state is identified. By locking the modal direction of the lower bound of the entropy gradient in the fuzzy membership value output by the fuzzy logic reasoning model, a neutral task buffer control band is constructed to shield the real-time control signal output of the feedforward neural network in the situation prediction module.

[0020] The coordination module is used to observe and provide feedback on the evolution of the industrial robot's real-time state input data within a preset time window in the neutral task buffer control zone, so as to extract dynamic response features that characterize the state evolution trajectory of the industrial robot, and judge the stabilization trend of the entropy deviation index value based on the extracted dynamic response features. When the stabilization trend recovers to the preset fluctuation range, the instantaneous control signal output of the situation prediction module is restored, thus forming coordinated control between the attribution module and the situation prediction module.

[0021] In a preferred embodiment, the situation prediction module is used to number the raw state input data obtained by continuous sampling in chronological order during the execution of the industrial robot, and construct an embedded state sequence matrix according to a preset sliding window length. Each row in the state sequence matrix represents the distribution trajectory of state parameters in adjacent time steps, and each column corresponds to the change dimension of a type of state parameter.

[0022] A time-coupled encoding operation is performed on the state sequence matrix, and a position offset nested function is used to calculate the partial derivatives of the state parameters at each time step with the change direction between them and their adjacent time steps, so as to construct a difference mapping tensor that characterizes the state evolution trend.

[0023] A feedforward neural network with at least three hidden layers is constructed. The hidden layers contain response control units. Each response control unit includes a set of response gate functions constructed based on the gradient directionality of the state parameters. The response gate functions are used to receive the difference mapping tensor as input and generate a compressed and encoded state trend signal at the output of the response gate functions.

[0024] The feedforward neural network is trained, and the trend residual function is used as the loss index during the training process. The trend residual function is used to evaluate the trend deviation between the state trend signal and the trajectory formed by the industrial robot in the real task path; it also includes using the backpropagation algorithm to optimize the connection weight parameters between the layers in the feedforward neural network.

[0025] After training is completed, the real-time sampled state input data is transformed into a state sequence matrix within the current time window, and a difference mapping tensor is generated according to the same steps. The difference mapping tensor is then input into the trained feedforward neural network, and the output includes the predicted output results of the task trend direction vector, modality offset probability distribution, and path transition rate.

[0026] The predicted output is input into the attribution determination module to provide support for predicting the task evolution path based on state trends, and to support the fuzzy logic reasoning model in real-time attribution determination of critical control states.

[0027] In a preferred embodiment, in the attribution determination module, for each control mode, a fuzzy membership function with adjustable shape parameters is constructed. The fuzzy membership function is constructed in the form of a piecewise nonlinear mapping. The membership start point, membership peak point and decay inflection point are set according to the state boundary conditions in the industrial robot task to describe the attribution transition characteristics of the state parameters in the modal boundary interval.

[0028] Based on the state membership expression requirements under multiple control modes, a rule base structure for a fuzzy logic reasoning model is designed, where each fuzzy rule is represented as a conditional association mapping between the input state parameter range and the corresponding modal label.

[0029] In a preferred embodiment, during the determination execution phase of the attribution module, the current state input data is arranged according to the preset input dimension order in the fuzzy rule base structure, and fuzzification encoding is performed on the state parameters of each input dimension; rule matching is performed sequentially on the fuzzification encoding results to mark all fuzzy rules that meet the preconditions; the activation value of each activated fuzzy rule is calculated, and its corresponding control mode label is bound to the activation value; by performing weighted summation on the activation values ​​of all activated rules under the same control mode, modal response coefficients corresponding to each control mode are generated, and the vector set formed by the modal response coefficients is the intermediate representation tensor of multimodal fusion inference;

[0030] Normalization is performed on the intermediate representation tensor formed by the modal response coefficients; based on the result of the normalization operation, a fuzzy membership value set containing all control modes is output, and the fuzzy membership value set serves as reference data for determining the distribution of the current state of the industrial robot among the control modes.

[0031] In a preferred embodiment, a collaborative discrimination model is constructed through a comparison module. The task trend prediction results output by the situation prediction module and the fuzzy membership values ​​output by the determination and attribution module are mapped one-to-one according to the control mode to a joint mode representation vector. The cosine value of the vector angle between the prediction channel of the situation prediction module and the fuzzy channel of the determination and attribution module and the normalized offset are calculated. The difference vector between the two in the control mode attribution dimension is extracted as the difference information of mode determination.

[0032] In a preferred embodiment, when the misalignment identification module calculates the control decision entropy deviation index value of the difference information, the entropy deviation index value is constructed based on the degree of deviation of each modal dimension in the difference vector to construct an information entropy distribution, and the standard deviation entropy value of the information entropy distribution is statistically analyzed.

[0033] The deviation degree includes the calculation operation based on the absolute difference between the predicted output result and the fuzzy membership value in each modal dimension. The difference between the predicted probability of the predicted output result and the corresponding fuzzy membership value under each control mode is obtained. Then, the difference of all modal dimensions is normalized and used as the offset weight distribution to construct the probability proportion of each modal dimension in the control decision entropy deviation index value.

[0034] The technical effects and advantages of this invention are as follows:

[0035] 1. This invention constructs a collaborative discrimination model of situation prediction and fuzzy reasoning, and introduces an entropy deviation index to identify critical state misalignment, thereby solving the control disorder problem caused by the lack of entropy weight coordination between prediction and fuzzy mechanisms.

[0036] 2. A sliding window is used to construct the state sequence matrix and extract the difference mapping tensor. Combined with the response gating function in the feedforward neural network, the continuity of task trend prediction and the forward response capability in complex dynamic scenarios are improved.

[0037] 3. By constructing a modal attribution distribution through fuzzy rule matching and response coefficient calculation, and combining it with normalization operation to output a set of fuzzy membership values, the distinguishability between multi-mode control modes and the compatibility of modal attribution expressions are enhanced.

[0038] 4. Based on the differences in modal judgment results, a control decision entropy deviation index is constructed. The system can identify risk states before conflicts occur, enter the buffer control zone in a timely manner, reduce the risk of task execution interruption, and improve overall stability.

[0039] 5. By continuously feeding back the state evolution trajectory and detecting the stabilization trend of the entropy deviation value, the system can adaptively recover the feedforward control path, realizing the closed-loop connection between fuzzy and predictive mechanisms. Attached Figure Description

[0040] Figure 1 This is a flowchart of the situation prediction process in the system of this invention.

[0041] Figure 2 This is a flowchart of the system's attribution determination process in this invention.

[0042] Figure 3 This is a flowchart of the modal comparison and misalignment identification process of the system in this invention.

[0043] Figure 4 This is a flowchart illustrating the construction process of the buffer control band in the system of this invention.

[0044] Figure 5 This is a flowchart of the coordinated control and recovery process of the system in this invention.

[0045] Figure 6 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Refer to the instruction manual appendix Figure 1-6 An embodiment of the present invention provides an adaptive control system for an industrial robot that integrates a situation prediction mechanism, comprising a situation prediction module, an attribution determination module, a comparison module, a misalignment identification module, a buffer module, and a coordination module.

[0048] The situation prediction module is used to continuously sample and acquire state input data during the execution of an industrial robot, and input the state input data into a feedforward neural network that has the ability to construct a time-series feature mapping, so as to obtain the predicted output result of the robot task trend.

[0049] The attribution determination module is used to input the state input data from the situation prediction module into the fuzzy logic reasoning model based on the fuzzy membership function, and to perform fuzzy membership reasoning on the attribution probability of the current state of the industrial robot among multiple control modes, so as to obtain the fuzzy membership value of the current state under each control mode.

[0050] The comparison module is used to build a collaborative discrimination model. It compares the prediction output obtained from the situation prediction module with the fuzzy membership value obtained from the determination and attribution module to extract the difference information between the two in the modality determination results.

[0051] The misalignment identification module is used to calculate the control decision entropy deviation index value of the difference information, determine whether the current state of the industrial robot is in a modal critical misalignment state, and when the entropy deviation index value exceeds the preset threshold, it is marked as a conflict risk state.

[0052] The buffer module is used to activate the fuzzy priority control mechanism when a conflict risk state is identified. By locking the modal direction of the lower bound of the entropy gradient in the fuzzy membership value output by the fuzzy logic reasoning model, a neutral task buffer control band is constructed to shield the real-time control signal output of the feedforward neural network in the situation prediction module.

[0053] The coordination module is used to observe and provide feedback on the evolution of the real-time state input data of the industrial robot within a preset time window in the neutral task buffer control zone, so as to extract the dynamic response characteristics that characterize the state evolution trajectory of the industrial robot, and judge the stabilization trend of the entropy deviation index value based on the extracted dynamic response characteristics. When the stabilization trend recovers to the preset fluctuation range, the instantaneous control signal output of the situation prediction module is restored, forming a coordinated control between the determination module and the situation prediction module to ensure the continuity and stability of the task response process.

[0054] The coordination module is used to perform the following operations within the neutral task buffer control band:

[0055] 1) State evolution observation: Acquire continuous state input data of the industrial robot within the current preset time window, and construct its state evolution trajectory within the window;

[0056] 2) Dynamic response feature extraction: Based on the state evolution trajectory, dynamic response features reflecting its changing trend, response directionality and disturbance stability are extracted;

[0057] 3) Entropy deviation stabilization judgment: Using the extracted dynamic response characteristics, calculate the evolution trend of the control decision entropy deviation index value, and determine whether the trend returns to the system's preset fluctuation stability range;

[0058] 4) Control signal recovery: When the stabilization trend of the entropy deviation index value meets the recovery conditions, the real-time control signal output of the situation prediction module is reactivated to form a coordinated control path with the attribution determination module to ensure continuous and stable task response.

[0059] Define the robot's state input data within the time window [t0, t0+T] as follows:

[0060] S(t) = {s1(t), s2(t), ..., s n (t)}, t∈[t0,t0+T]

[0061] The dynamic response feature vector for constructing the state evolution trajectory is:

[0062]

[0063] Based on the dynamic response feature vector, the stabilization trend function for calculating the control decision entropy deviation index value is as follows:

[0064]

[0065] The output control signal will resume marking when the following conditions are met:

[0066] Ξ(t0)≤ε th

[0067] Where S(t) is the multidimensional state input data sequence of the industrial robot at time t, which includes parameters such as position, velocity, attitude, and force in practical applications; n is the total number of dimensions of the state input data; T is the time window period; Φ(·) is the dynamic response feature constructor, whose inputs include the first derivative (velocity), the second derivative (acceleration), and the attitude transformation gradient Δθ(t), and the output is used to measure the intensity and directionality of the current state evolution; F(t0) represents the dynamic response feature vector extracted at the start of the time window t0; P mod (t) represents the modality attribution difference vector (the probabilistic basis for entropy bias calculation) constructed between the feedforward prediction result and the fuzzy membership value at each time t. The distribution entropy calculation function represents the control decision entropy deviation index value, used to calculate the entropy value of different modal attribution deviations; Ω(F(t)) is the response weighting function, used to adjust the influence of the entropy value at each time point on the overall trend according to the dynamic response characteristics; Λ(·) is the stabilization trend aggregation function, used to perform integrated evaluation on the weighted entropy evolution within the entire time window; Ξ(t0) represents the trend function of the control decision entropy deviation index value in the current time window, used to determine whether the recovery condition has been met; ε th The upper limit threshold for the stability of entropy deviation trend fluctuations is preset by the system.

[0068] The situation prediction module is used to number the raw state input data obtained by continuous sampling in chronological order during the execution of the industrial robot, and construct an embedded state sequence matrix according to a preset sliding window length. Each row in the state sequence matrix represents the distribution trajectory of state parameters in adjacent time steps, and each column corresponds to the change dimension of a type of state parameter.

[0069] A time-coupled encoding operation is performed on the state sequence matrix, and a position offset nested function is used to calculate the partial derivatives of the state parameters at each time step with the change direction between them and their adjacent time steps, so as to construct a difference mapping tensor that characterizes the state evolution trend.

[0070] A feedforward neural network with at least three hidden layers is constructed. The hidden layers contain response control units. Each response control unit includes a set of response gate functions constructed based on the gradient directionality of the state parameters. The response gate functions are used to receive the difference mapping tensor as input and generate a compressed and encoded state trend signal at the output of the response gate functions.

[0071] The feedforward neural network is trained, and the trend residual function is used as the loss index during the training process. The trend residual function is used to evaluate the trend deviation between the state trend signal and the trajectory formed by the industrial robot in the real task path; it also includes using the backpropagation algorithm to optimize the connection weight parameters between the layers in the feedforward neural network.

[0072] After training is completed, the real-time sampled state input data is transformed into a state sequence matrix within the current time window, and a difference mapping tensor is generated according to the same steps. The difference mapping tensor is then input into the trained feedforward neural network, and the output includes the predicted output results of the task trend direction vector, modality offset probability distribution, and path transition rate.

[0073] The predicted output is input into the attribution determination module to provide support for predicting the task evolution path based on state trends, and to support the fuzzy logic reasoning model in real-time attribution determination of critical control states.

[0074] Specifically, the differential mapping tensor representation is generated by following the same steps, using the methods employed in the training phase: constructing an embedded state sequence matrix with a preset sliding window length, performing time-coupled encoding on the state sequence matrix, and calculating partial derivatives using a nested position offset function. This process generates a differential mapping tensor that represents the evolution trend of the current state.

[0075] In the attribution determination module, for each control mode, a fuzzy membership function with adjustable shape parameters is constructed. The fuzzy membership function is constructed in the form of a piecewise nonlinear mapping. The membership start point, membership peak point, and decay inflection point are set according to the common state boundary conditions in industrial robot tasks to describe the attribution transition characteristics of state parameters in the modal boundary interval. The common state boundary conditions include robot position limits, velocity mutation points, posture stability boundaries, load critical values, environmental disturbance thresholds, and control mode switching regions.

[0076] Based on the state membership representation requirements under multiple control modes, a rule base structure for a fuzzy logic reasoning model is designed. Each fuzzy rule is represented as a conditional association mapping between the input state parameter range and the corresponding modal label. Modal conflict constraint coefficients are introduced to adjust the reasoning priority weight of conflicting modes when multiple rules are activated simultaneously.

[0077] During the determination execution phase of the attribution module, the current state input data is arranged according to the preset input dimension order in the fuzzy rule base structure, and fuzzification encoding is performed on the state parameters of each input dimension; rule matching is performed on the fuzzification encoding results in sequence to mark all fuzzy rules that meet the preconditions; the activation value of each activated fuzzy rule is calculated, and its corresponding control mode label is bound to the activation value; by performing weighted summation on the activation values ​​of all activated rules under the same control mode, modal response coefficients corresponding to each control mode are generated, and the vector set formed by the modal response coefficients is the intermediate representation tensor of multimodal fusion inference;

[0078] A normalization operation is performed on the intermediate representation tensor composed of modal response coefficients. During the normalization process, a distribution consistency constraint factor is introduced to maintain the separation degree of response value intervals between each control mode. Based on the result of the normalization operation, a fuzzy membership value set containing all control modes is output. The fuzzy membership value set serves as reference data for determining the distribution of the current state of the industrial robot among the control modes.

[0079] A collaborative discrimination model is constructed by comparing modules. The task trend prediction results output by the situation prediction module and the fuzzy membership values ​​output by the determination and attribution module are mapped one-to-one according to the control mode to a joint mode representation vector. The cosine value of the vector angle between the prediction channel of the situation prediction module and the fuzzy channel of the determination and attribution module and the normalized offset are calculated. The difference vector between the two in the control mode attribution dimension is extracted as the difference information of mode determination.

[0080] When the misalignment identification module calculates the control decision entropy deviation index value of the difference information, the entropy deviation index value is constructed based on the degree of deviation of each modal dimension in the difference vector to construct an information entropy distribution, and the standard deviation entropy value of the information entropy distribution is statistically analyzed.

[0081] The deviation degree includes the calculation operation based on the absolute difference between the predicted output result and the fuzzy membership value in each modal dimension. The difference between the predicted probability of the predicted output result and the corresponding fuzzy membership value under each control mode is obtained. Then, the difference of all modal dimensions is normalized and used as the offset weight distribution to construct the probability proportion of each modal dimension in the control decision entropy deviation index value.

[0082] It should be noted that this solution addresses the core issue of "control splitting and response defocusing caused by the lack of dynamic entropy coordination mechanism between the situation prediction mechanism and the fuzzy logic reasoning mechanism during the execution of multimodal tasks by industrial robots." It constructs an industrial robot adaptive control system that integrates the situation prediction mechanism. Its workflow is based on six clearly structured functional modules that unfold sequentially: situation prediction module, attribution determination module, comparison module, misalignment identification module, buffer module, and coordination module.

[0083] During the industrial robot's task execution, the system continuously samples the state input data through the situation prediction module and encodes it into a state sequence matrix with temporal relationships. By setting a sliding window structure, this matrix can retain the historical evolution trajectory of the state parameters. Based on this, a time-coupled coding mechanism is introduced. The gradient direction of each time step state at the preceding and following time nodes is calculated using a nested position offset function to obtain a difference mapping tensor representing the state evolution trend. Subsequently, a multi-layer feedforward neural network with embedded response control units is constructed, and the trend signal is compressed and encoded through a response gating function. The entire training process uses the trend residual function as the loss index and optimizes the connection weights through the backpropagation algorithm to improve prediction accuracy.

[0084] The attribution determination module processes the same state input data in parallel from another path. The system constructs a fuzzy rule base based on the fuzzy logic reasoning mechanism and maps the state parameters to the membership space corresponding to each control mode through the fuzzy membership function. Specifically, the attribution determination module performs fuzzification encoding to fuzzify the input data, and then outputs the modal response coefficient matrix under the control mode through rule matching and activation calculation, forming the inference intermediate tensor of the fusion expression. Subsequently, a normalization operation with distribution consistency constraints is performed to generate a set of fuzzy membership values ​​that meet the requirements of modal discriminability and compatibility, which is used to characterize the probability distribution of the current state's attribution among the control modes.

[0085] The comparison module structurally compares the state trend prediction results output by the situation prediction module with the fuzzy membership values ​​generated by the attribution determination module, calculating their differences in the modal dimension. The misalignment identification module further models these differences using decision entropy, generating a control decision entropy deviation index value based on the degree of modal offset between the predicted output and the fuzzy inference output. This entropy deviation index value reflects whether the current state is in a modal critical misalignment state, i.e., an irreconcilable conflict has occurred between the predicted trend and the fuzzy attribution. When the entropy deviation index value exceeds a preset threshold, the system will identify the state as a conflict risk state and activate the buffer mechanism.

[0086] In the buffer module, the system automatically pauses the feedforward network's direct control output on the state and instead establishes a neutral task buffer control zone based on the lower bound direction of the entropy gradient output in the fuzzy logic reasoning model. This buffer control zone can be regarded as a fuzzy priority scheduling interval, which temporarily maintains the robot's operation in the intermediate state between modes, thereby preventing premature or incorrect mode switching due to misjudgment.

[0087] The coordination module continuously monitors the dynamic response changes of the state input data within a set time window within the buffer control zone, extracts the state evolution trajectory, and judges whether the control decision entropy deviation index value has stabilized based on its trend change. When the index value falls back to the preset fluctuation range, the system is considered to have recovered modal consistency, that is, it is judged that the feedforward prediction and fuzzy inference have reached modal consensus. At this time, the feedforward neural network control output is reactivated, and the control power is transferred from the fuzzy logic mechanism back to the feedforward prediction mechanism, forming a smooth connection of task response and ensuring the response stability and task continuity of the entire control system.

[0088] The above structural design is adopted because existing industrial robot control systems cannot identify modal critical misalignment states in their situation prediction mechanisms. At the same time, although fuzzy logic reasoning models can maintain control continuity within fuzzy boundaries, they cannot provide trend prediction capabilities. Therefore, this system solves the disconnect between feedforward neural networks and fuzzy reasoning by heterogeneously coordinating the information sources of the two and establishing a transition judgment mechanism between decision domains through entropy indices. This improves the robustness and intelligence of the control system in multimodal dynamic switching tasks.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive control system for an industrial robot integrating a situation prediction mechanism, comprising a situation prediction module, an attribution determination module, a comparison module, a misalignment identification module, a buffer module, and a coordination module, characterized in that: The situation prediction module is used to continuously sample and acquire state input data during the execution of an industrial robot, and input the state input data into a feedforward neural network that has the ability to construct a time-series feature mapping, so as to obtain the predicted output result of the robot task trend. The attribution determination module is used to input the state input data from the situation prediction module into the fuzzy logic reasoning model based on the fuzzy membership function, and to perform fuzzy membership reasoning on the attribution probability of the current state of the industrial robot among multiple control modes, so as to obtain the fuzzy membership value of the current state under each control mode. The comparison module is used to build a collaborative discrimination model. It compares the prediction output obtained from the situation prediction module with the fuzzy membership value obtained from the determination and attribution module to extract the difference information between the two in the modality determination results. The misalignment identification module is used to calculate the control decision entropy deviation index value of the difference information, determine whether the current state of the industrial robot is in a modal critical misalignment state, and when the entropy deviation index value exceeds the preset threshold, it is marked as a conflict risk state. The situation prediction module is used to number the raw state input data obtained by continuous sampling in chronological order during the execution of the industrial robot, and construct an embedded state sequence matrix according to a preset sliding window length. Each row in the state sequence matrix represents the distribution trajectory of state parameters in adjacent time steps, and each column corresponds to the change dimension of a type of state parameter. A time-coupled encoding operation is performed on the state sequence matrix, and a position offset nested function is used to calculate the partial derivatives of the state parameters at each time step with the change direction between them and their adjacent time steps, so as to construct a difference mapping tensor that characterizes the state evolution trend. A feedforward neural network with at least three hidden layers is constructed. The hidden layers contain response control units. Each response control unit includes a set of response gate functions constructed based on the gradient directionality of the state parameters. The response gate functions are used to receive the difference mapping tensor as input and generate a compressed and encoded state trend signal at the output of the response gate functions. The feedforward neural network is trained, and the trend residual function is used as the loss index during the training process. The trend residual function is used to evaluate the trend deviation between the state trend signal and the trajectory formed by the industrial robot in the real task path; it also includes using the backpropagation algorithm to optimize the connection weight parameters between the layers in the feedforward neural network. After training is completed, the real-time sampled state input data is transformed into a state sequence matrix within the current time window, and a difference mapping tensor is generated according to the same steps. The difference mapping tensor is then input into the trained feedforward neural network, and the output includes the predicted output results of the task trend direction vector, modality offset probability distribution, and path transition rate. The predicted output is input into the attribution determination module to provide support for predicting the task evolution path based on state trends, and to support the fuzzy logic reasoning model in real-time attribution determination of critical control states.

2. The industrial robot adaptive control system with integrated situation prediction mechanism according to claim 1, characterized in that: The buffer module is used to activate the fuzzy priority control mechanism when a conflict risk state is identified. By locking the modal direction of the lower bound of the entropy gradient in the fuzzy membership value output by the fuzzy logic reasoning model, a neutral task buffer control band is constructed to shield the real-time control signal output of the feedforward neural network in the situation prediction module. The coordination module is used to observe and provide feedback on the evolution of the industrial robot's real-time state input data within a preset time window in the neutral task buffer control zone, so as to extract dynamic response features that characterize the state evolution trajectory of the industrial robot, and judge the stabilization trend of the entropy deviation index value based on the extracted dynamic response features. When the stabilization trend recovers to the preset fluctuation range, the instantaneous control signal output of the situation prediction module is restored, thus forming coordinated control between the attribution module and the situation prediction module.

3. The industrial robot adaptive control system with integrated situation prediction mechanism according to claim 2, characterized in that: In the determination and attribution module, for each control mode, a fuzzy membership function with adjustable shape parameters is constructed. The fuzzy membership function is constructed in the form of piecewise nonlinear mapping. The membership start point, membership peak point and decay inflection point are set according to the state boundary conditions in the industrial robot task to describe the attribution transition characteristics of the state parameters in the modal boundary interval. Based on the state membership expression requirements under multiple control modes, a rule base structure for a fuzzy logic reasoning model is designed, where each fuzzy rule is represented as a conditional association mapping between the input state parameter range and the corresponding modal label.

4. The industrial robot adaptive control system with integrated situation prediction mechanism according to claim 3, characterized in that: During the determination execution phase of the attribution module, the current state input data is arranged according to the preset input dimension order in the fuzzy rule base structure, and fuzzification encoding is performed on the state parameters of each input dimension; rule matching is performed on the fuzzification encoding results in sequence to mark all fuzzy rules that meet the preconditions; the activation value of each activated fuzzy rule is calculated, and its corresponding control mode label is bound to the activation value; by performing weighted summation on the activation values ​​of all activated rules under the same control mode, modal response coefficients corresponding to each control mode are generated, and the vector set formed by the modal response coefficients is the intermediate representation tensor of multimodal fusion inference; Normalization is performed on the intermediate representation tensor formed by the modal response coefficients; based on the result of the normalization operation, a fuzzy membership value set containing all control modes is output, and the fuzzy membership value set serves as reference data for determining the distribution of the current state of the industrial robot among the control modes.

5. An industrial robot adaptive control system with an integrated situation prediction mechanism according to claim 4, characterized in that: A collaborative discrimination model is constructed by comparing modules. The task trend prediction results output by the situation prediction module and the fuzzy membership values ​​output by the determination and attribution module are mapped one-to-one according to the control mode to a joint mode representation vector. The cosine value of the vector angle between the prediction channel of the situation prediction module and the fuzzy channel of the determination and attribution module and the normalized offset are calculated. The difference vector between the two in the control mode attribution dimension is extracted as the difference information of mode determination.

6. An industrial robot adaptive control system with an integrated situation prediction mechanism according to claim 5, characterized in that: When the misalignment identification module calculates the control decision entropy deviation index value of the difference information, the entropy deviation index value is constructed based on the degree of deviation of each modal dimension in the difference vector to construct an information entropy distribution, and the standard deviation entropy value of the information entropy distribution is statistically analyzed. The deviation degree includes the calculation operation based on the absolute difference between the predicted output result and the fuzzy membership value in each modal dimension. The difference between the predicted probability of the predicted output result and the corresponding fuzzy membership value under each control mode is obtained. Then, the difference of all modal dimensions is normalized and used as the offset weight distribution to construct the probability proportion of each modal dimension in the control decision entropy deviation index value.

Images