Adaptive control method and system based on multi-parameter perception

Abstract

The invention provides a self-adaptive control method and system based on multi-parameter perception, and relates to the technical field of intelligent control, and the method comprises the steps: obtaining sensing measurement data and unstructured object description information, and forming an observation sequence containing sensing measurement data time sequence segments and associated object description information thereof; generating model input data based on the observation sequence, inputting the model input data into a pre-training semantic inference model to obtain object semantic output, and determining a semantic tag set and a credibility set according to the object semantic output; and determining a control prior parameter set based on the semantic tag set and the credibility set, and further generating a control vector based on the control prior parameter set and the observation sequence. Robust processing of uncertainty of object attributes is realized by introducing semantic tags and credibility constraints, strategy deviation and instability caused by misjudgment of the object attributes can be suppressed under the condition that multi-source sensing data fluctuates along with time, and consistency and robustness of control results are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control, in particular to a self-adaptive control method and system based on multi-parameter perception. BACKGROUND

[0002] The intelligent clothes drying machine is usually configured with temperature and humidity, wind speed, light, weight or current sensors, and drives the actuator in combination with a preset control strategy to achieve a specific operation target. In related technologies, the selection of control strategy parameters often depends on the judgment of the controlled object attributes and operating conditions, such as threshold judgment based on sensor measurement data, mode selection based on user input, or parameter mapping based on an experience rule library.

[0003] However, the clothes drying machine running process has significant time-varying and multi-source inconsistency: on the one hand, multi-source perception data fluctuates over time and is easily affected by external disturbances such as window ventilation, environmental temperature mutation, load change, local shielding, etc., resulting in non-stationary changes; on the other hand, the attributes of the controlled object cannot be directly observed, and the object description information provided by the user may be ambiguous, default or inconsistent in caliber, making the judgment of object attributes uncertain.

[0004] In the above case, if the control system misjudges the object attributes, it is easy to cause the control parameter selection to deviate from the actual working condition, making the driving instruction unstable in aspects such as lifting, air volume, energy consumption or operation rhythm, and further causing control effect degradation or oscillation, and it may be difficult to recover in the subsequent period by continuously using the inappropriate parameters.

[0005] Therefore, the prior art needs to solve the problem of how to suppress the strategy deviation and instability caused by the misjudgment of object attributes in the control parameter determination and instruction output process when multi-source perception data changes over time and object attributes are difficult to determine. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a self-adaptive control method and system based on multi-parameter perception.

[0007] In a first aspect, the present application provides a self-adaptive control method based on multi-parameter perception, comprising:

[0008] Obtaining sensor measurement data and user-provided unstructured object description information to form an observation sequence;

[0009] Based on the observation sequence, generating model input data, inputting the model input data into a pre-trained semantic inference model to obtain object semantic output, determining a semantic tag set and a confidence set corresponding one-to-one to the semantic tag set according to the object semantic output;

[0010] Based on the semantic tag set and the credibility set, a control prior parameter set is determined according to a preset mapping rule, a control vector is generated based on the control prior parameter set and the observation sequence, and the control vector is converted into a driving instruction signal output.

[0011] Optionally, the observation sequence includes a time sequence segment of the sensor measurement data and unstructured object description information associated with the time sequence segment.

[0012] Optionally, the observation sequence includes a time sequence segment of the sensor measurement data, unstructured object description information associated with the time sequence segment, and a physical state sequence, the physical state sequence being a physical characterization of the time sequence segment, at least one physical state quantity being obtained by performing a physical state transformation on the time sequence segment.

[0013] Optionally, the generation model input data includes:

[0014] Performing quantization encoding on the observation sequence generates a time sequence label sequence;

[0015] Obtaining a running phase identifier determined by a preset running state machine;

[0016] Based on the time sequence label sequence, the running phase identifier, and the unstructured object description information, the model input data is constructed.

[0017] Optionally, the generation model input data further includes:

[0018] Obtaining an actuator control instruction sequence of the previous control cycle, encoding the actuator control instruction sequence to generate a control label sequence, and adding the control label sequence to the model input data.

[0019] Optionally, the pre-trained semantic inference model includes a plurality of semantic adapter parameter groups, a target semantic adapter parameter group is selected based on the model input data, and the object semantic output is generated under the action of the target semantic adapter parameter group.

[0020] Optionally, the performing physical state transformation to obtain a physical state sequence includes:

[0021] Performing physical state transformation on the sensor measurement data at each sampling time to obtain a candidate physical state vector, and calculating a physical consistency residual based on the compatibility relationship between at least two types of redundant physical state quantities in the candidate physical state vector;

[0022] The candidate physical state vector is divided into a steady state point and a perturbation point according to the physical consistency residual, and the perturbation point is segmented and reconstructed with the steady state point after projection correction under preset physical feasible region constraints to obtain the physical state sequence.

[0023] Optionally, the segmented reconstruction comprises:

[0024] A segment boundary formed by adjacent switching of the steady state point and the perturbation point is determined, and based on the physical state vector and a preset reference process range, a first process feature representing a deviation driving force of the physical state vector relative to the reference process range and a second process feature representing a degree of advancement of the physical state vector along the reference process range are determined;

[0025] A boundary continuity constraint is imposed on the reconstructed physical state vector at the segment boundary, the boundary continuity constraint comprising: the second process feature is continuous at the segment boundary and does not appear a change opposite to a preset change direction along the sampling time sequence, the preset change direction being one of an increasing direction or a decreasing direction, and a jump amount of the first process feature at the segment boundary is constrained within a reference range defined by the preset reference process range; and

[0026] Arc length parameterization is performed on the reconstruction result satisfying the boundary continuity constraint in a physical state space constituted by at least two types of physical state quantities in the physical state vector, and resampling is performed at an equal arc length step to obtain the physical state sequence.

[0027] Optionally, further comprising:

[0028] The sensing measurement data is updated within a preset control period after the control vector is output, and the updated sensing measurement data is subjected to the physical state transformation to obtain an updated physical state sequence;

[0029] The first process feature and the second process feature are determined based on the updated physical state sequence, and after the first process feature and the second process feature are re-parameterized to arc length coordinates corresponding to the reference process range in the physical state space, a deviation degree of the first process feature and the second process feature relative to the reference process range is calculated, and a prior consistency index is determined based on the deviation degree;

[0030] In response to the prior consistency index not satisfying a preset consistency condition, at least two semantic labels are selected based on the credibility set and label weights are respectively determined, the control prior parameters corresponding to the at least two semantic labels are weighted and fused according to the preset mapping rule to generate a mixed control prior parameter set, and a subsequent control vector is generated based on the mixed control prior parameter set;

[0031] The label weight is determined according to the set of confidence degrees, the physical consistency residual, and the deviation degree.

[0032] Optionally, the reference process range includes a first reference trajectory band set for the first process feature and a second reference trajectory band set for the second process feature; the first reference trajectory and the second reference trajectory share an arc length coordinate corresponding to the reference process range;

[0033] The first reference trajectory band is defined by a first upper boundary trajectory and a first lower boundary trajectory under the arc length coordinate of the first reference trajectory, and the second reference trajectory band is defined by a second upper boundary trajectory and a second lower boundary trajectory under the arc length coordinate of the second reference trajectory.

[0034] The calculation of the deviation degree includes: taking the first process feature re-parameterized to the arc length coordinate corresponding to the first reference trajectory band as a first process feature trajectory, and determining whether the first process feature trajectory falls into the first reference trajectory band at each arc length position;

[0035] Taking the second process feature re-parameterized to the arc length coordinate corresponding to the second reference trajectory band as a second process feature trajectory, and determining whether the second process feature trajectory falls into the second reference trajectory band at each arc length position; for any process feature trajectory exceeding the arc length position of the corresponding reference trajectory band, determining an out-of-limit amplitude and accumulating the deviation degree according to the arc length.

[0036] In a second aspect, the present application provides a self-adaptive control system based on multi-parameter perception, comprising:

[0037] A collection module is configured to acquire sensing measurement data and unstructured object description information provided by a user, and form an observation sequence.

[0038] A processing module is configured to generate model input data based on the observation sequence, input the model input data into a pre-trained semantic inference model, obtain object semantic output, determine a set of semantic labels and a set of confidence degrees corresponding to the set of semantic labels according to the object semantic output.

[0039] An output module is configured to determine a set of control prior parameters according to a preset mapping rule based on the set of semantic labels and the set of confidence degrees, generate a control vector based on the set of control prior parameters and the observation sequence, and convert the control vector into a driving instruction signal output.

[0040] Compared with existing technologies, the beneficial effects achieved by this application are as follows: By associating time-series segments of sensor measurement data with user-provided object description information, an observation sequence is constructed and model input data is generated. A pre-trained semantic inference model is used to output object semantics and form semantic labels and a set of confidence levels. This allows the determination of control parameters to no longer rely solely on a single threshold or fixed rules, thus providing a decision-making basis with confidence constraints even when object attributes are difficult to determine accurately. Furthermore, based on the set of semantic labels and the set of confidence levels, a set of prior control parameters is obtained according to mapping rules. This, combined with the observation sequence, generates a control vector and outputs it as a driving command signal. This helps to achieve more robust parameter selection and command generation under time-varying fluctuations in multi-source sensing data, reduces the amplification effect of object attribute misjudgment on the control strategy, improves the adaptability of the control process to disturbances and uncertain inputs, and thus enhances the consistency and robustness of the control results, reducing the risk of control performance degradation and instability. Attached Figure Description

[0041] Figure 1 A flowchart illustrating an adaptive control method based on multi-parameter sensing, provided as an embodiment of this application;

[0042] Figure 2 A flowchart illustrating a method for generating model input data provided in an embodiment of this application;

[0043] Figure 3 A flowchart illustrating a method for obtaining a physical state sequence provided in this application embodiment;

[0044] Figure 4 This is a schematic diagram of an adaptive control system based on multi-parameter sensing, provided as an embodiment of this application. Detailed Implementation

[0045] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0046] Example 1

[0047] See Figure 1 This application provides an adaptive control method based on multi-parameter sensing, including steps S101 to S103, wherein:

[0048] S101: Acquire sensor measurement data and unstructured object description information provided by the user to form an observation sequence;

[0049] S102: Based on the observation sequence, generate model input data, input the model input data into the pre-trained semantic inference model to obtain object semantic output, and determine the semantic label set and the credibility set corresponding to the semantic label set according to the object semantic output.

[0050] S103: Based on the semantic tag set and the confidence set, determine the control prior parameter set according to the preset mapping rule, generate a control vector based on the control prior parameter set and the observation sequence, and convert the control vector into a driving command signal output.

[0051] Regarding the above S101:

[0052] The observation sequence may include two aspects of information: time-series segments of the sensing measurement data and unstructured object description information associated with the time-series segments.

[0053] In some high-configuration embodiments, the observation sequence may include three aspects of information: a time-series segment of the sensing measurement data, unstructured object description information associated with the time-series segment, and a physical state sequence; wherein, the physical state sequence is the physical representation of the time-series segment, and at least one physical state quantity is obtained by performing a physical state transformation on the time-series segment.

[0054] In one embodiment, the controller of the smart clothes drying rack acquires multi-source sensor measurement data during operation and receives unstructured object description information input by the user to form an observation sequence for subsequent inference and control.

[0055] Sensing measurement data can be acquired by sensors or detection units that are electrically or communicatively connected to the controller. It is used to characterize the operating environment and / or equipment operating status. The sensing measurement data can be single-channel or multi-channel time series data, with sampling time or relative time identifiers.

[0056] Unstructured object description information is input by the user through a human-computer interaction interface. The human-computer interaction interface may include the buttons or touch interface of the clothes drying rack itself, mobile terminal applications, voice input to text, etc. Unstructured object description information is natural language or short text without a fixed format, used to describe the attributes of the controlled object, user intent or operational requirements. The description information may contain colloquial expressions, default information or synonyms and does not require pre-structured into fixed fields.

[0057] To form an observation sequence, the controller organizes the sensor measurement data chronologically and extracts a continuous set of samples from the sensor measurement data under preset window rules as time-series segments of the sensor measurement data. The window rules can be fixed-length windows, sliding windows, or event-triggered windows; the window length, step size, and trigger conditions are preset or configurable. The controller further establishes the association between unstructured object description information and time-series segments, enabling the unstructured object description information to be bound to its corresponding runtime segment. The association can be established based on timestamp alignment and / or session identifiers: for example, a running task, a startup cycle, or a user confirmation process can be considered as a session, and the unstructured object description information generated within that session can be bound to one or more time-series segments within that session; or, the generation time of the unstructured object description information can be aligned with the time range of the time-series segment, and the associated time-series segments can be determined according to rules of inclusion, nearest neighbor, or preset tolerance range.

[0058] Based on the above, the observation sequence consists of multiple observation units ordered by time. Each observation unit includes at least the sensor measurement data representation of the corresponding time segment and the unstructured object description information or its identifier associated with the time segment. This allows the observation sequence to simultaneously characterize the measurement state that changes over time and its corresponding object description semantics, providing a consistent data foundation for subsequent model input data construction, semantic inference, and control vector generation based on the observation sequence.

[0059] Regarding S102 above:

[0060] In one embodiment, after obtaining an observation sequence consisting of temporal segments of sensor measurement data and unstructured object description information, the controller generates model input data for semantic inference based on the observation sequence. The model input data is used to characterize the joint information of object description semantics and sensor temporal state. Its generation process may include: performing text normalization processing on the unstructured object description information, converting the text-normalized object description information into a text-labeled sequence; performing serialization representation on the temporal segments consistent with the sampling order, converting the temporal segments into a temporal-labeled sequence or a temporal-feature-labeled sequence; and combining the text-labeled sequence and the temporal-labeled sequence to form the model input data while maintaining the correlation between them.

[0061] Understandably, to meet the real-time requirements of edge inference, the combination can be implemented by splicing, aligning or packing, and the input length can be truncated or padded so that the model input data meets the input interface constraints of the pre-trained semantic inference model.

[0062] In one embodiment, the pre-trained semantic inference model is an offline pre-trained semantic inference model with fixed parameters. During the inference phase, the controller loads the model parameters and performs forward computation to output object semantic output. The object semantic output is used to characterize the semantic attribute inference result of the object corresponding to the observation sequence. The object semantic output may include a label score vector, label probability vector, or equivalent structured output information for a preset semantic label space. Based on the object semantic output, the controller determines a set of semantic labels according to preset label selection rules. The label selection rules may include threshold screening rules, Top-K selection rules, or a combination of both, so as to allow multiple candidate semantic labels to be output when there is uncertainty in the object attributes. The controller further determines a set of confidence levels that corresponds one-to-one with the set of semantic labels based on the object semantic output. The set of confidence levels is used to characterize the confidence level of each semantic label. In an optional implementation, the set of confidence levels is obtained by normalizing and calibrating the scores or probabilities in the object semantic output, or directly given by the confidence estimation components of the model output, so that the set of confidence levels can serve as a reliable basis for the subsequent determination of prior parameters of the control.

[0063] In one embodiment, to ensure that the semantic label set and the confidence set have a consistent interpretation within the control cycle, the controller can limit the semantic label set to label items from a preset label dictionary and limit the confidence set to numerical values ​​that correspond one-to-one with the label items. The range of values ​​and their semantic meanings can be pre-configured and fixed in the controller parameters to ensure consistency in the comparison and use of semantic inference results under different observation sequence conditions, thereby providing a stable semantic basis for subsequent control strategy parameter selection and drive instruction generation.

[0064] For example, the controller can generate an association identifier for each control cycle and use this association identifier to bind the set of sensor measurement data samples collected within the control cycle to the received object description information. Specifically, the controller can divide the continuously sampled sensor measurement data into several time segments according to a preset control cycle, with each time segment corresponding to a sampling time range. When object description information is received within this sampling time range, the controller associates the object description information with the corresponding time segment and writes it into the observation sequence. When multiple object description information messages are received within the same sampling time range, the latest message can be selected based on the timestamp, or multiple messages can be concatenated and merged according to a preset merging rule while retaining the original separator markers, to maintain a consistent correspondence between the text source and the time segment. This allows the subsequent model input data construction stage to determine the corresponding window between the text marker sequence and the time marker sequence, avoiding semantic inference ambiguities caused by mismatch between text semantics and time state.

[0065] For example, the controller can add segment tags and position tags to the generated text tag sequence and temporal tag sequence to distinguish modalities. Specifically, the text tag sequence can be obtained by word segmentation or sub-word segmentation and mapped to discrete tag numbers; the temporal tag sequence can be obtained by quantizing and encoding the multidimensional sensing measurement data at each sampling time, or by encoding the segment-level statistical features extracted from the temporal segments; the controller can set segment tags for the text tag sequence and the temporal tag sequence respectively, and set position tags or time position codes for them respectively, consistent with the sampling order, so that the model can distinguish the text semantic dimension and the temporal state dimension and maintain the temporal information.

[0066] Furthermore, when combining data to form model input data, preset separators can be inserted between text marker sequences and time sequence marker sequences, and boundary markers can be added to indicate control cycle boundaries. When the input length exceeds the maximum length allowed by the model interface, earlier time sequence samples or redundant text fragments can be truncated according to preset priorities. When the input length is insufficient, padding can be performed according to preset padding markers and corresponding attention masks can be generated to ensure the determinism and real-time performance of edge inference.

[0067] For example, a pre-trained semantic inference model may include a text encoding submodule for processing text-tagged sequences, a temporal encoding submodule for processing temporal-tagged sequences, a fusion submodule for fusing the two types of encoding results, and an output submodule for outputting object semantic output. The text encoding submodule can perform contextual semantic representation on the text-tagged sequences, and the temporal encoding submodule can perform state evolution representation on the temporal-tagged sequences; the fusion submodule can use any of the following methods—concatenation fusion, gated fusion, or cross-attention fusion—to fuse the text semantic representation and the temporal state representation into a joint representation; the output submodule can generate object semantic output based on the joint representation, and at least output a label score or label probability for a preset semantic label space.

[0068] In addition, to support the generation of the confidence set, the output submodule can further output confidence estimation components or uncertainty estimation components; in an optional implementation, the controller can perform calibration processing on the label score or label probability under the action of offline calibration parameters, so that the confidence set has consistent numerical semantics under different operating stages and different observation sequence conditions.

[0069] For example, the controller can encapsulate the semantic tag set and the confidence set into a semantic inference result structure and output it to the control prior parameter determination module. The semantic inference result structure includes at least: a semantic tag identifier list, a confidence list corresponding one-to-one with the semantic tag identifier list, and a control period identifier or time segment identifier corresponding to the semantic inference result.

[0070] The control prior parameter determination module can process the semantic inference result structure according to preset mapping rules. The preset mapping rules may include label filtering rules based on confidence thresholds, weighted fusion rules based on confidence, or a combination of the two, to generate a control prior parameter set. The control prior parameter set may consist of several parameter items, which may include target parameter values, allowable variation ranges, or parameter weights used to generate control vectors, so that the semantic inference results can enter the control parameter determination and drive instruction generation process in a structured and traceable manner.

[0071] Regarding the above S103:

[0072] In one embodiment, after obtaining the set of semantic tags and the set of confidence levels corresponding to the set of semantic tags, the controller calls a preset mapping rule to generate a set of control prior parameters, and generates a control vector based on the set of control prior parameters and the observation sequence, and then converts the control vector into a drive command signal output.

[0073] Among them, the preset mapping rules are used to establish the correspondence between semantic labels and control prior parameter templates. The control prior parameter set is used to provide prior constraints on the reference scope, weight configuration and constraint boundary of control calculation, so that the controller can still generate stable control output when there is uncertainty in the object attributes.

[0074] In an optional implementation, the preset mapping rules are implemented by a mapping table, a rule base, or a parameter template dictionary, and are stored in the parameter area or policy library of the controller. Each semantic tag corresponds to a set of control prior parameter templates. The control prior parameter templates include at least parameter items for characterizing the reference target or reference process range, parameter items for characterizing the control calculation weight or gain caliber, and constraint parameter items for characterizing the output limit and rate of change limit, so that the control prior parameter set can be used as the prior input of the control vector generation process.

[0075] In one embodiment, the controller determines the selection or fusion method of the control prior parameter set based on the confidence set, so that the confidence set forms a closed-loop application in the control link. Specifically, when the semantic tag set meets the single tag triggering condition, the controller selects the target semantic tag with the highest priority whose confidence meets the preset threshold, and looks up the control prior parameter template corresponding to the target semantic tag according to the mapping table, and instantiates the control prior parameter template into a control prior parameter set; when the semantic tag set meets the multi-tag triggering condition, the controller selects at least two semantic tags from the semantic tag set, and determines the parameter template corresponding to each semantic tag, and then determines the fusion weight based on the confidence set, performs weighted fusion or confidence threshold selection on the same type of parameter items in each parameter template, and obtains a mixed control prior parameter set.

[0076] Understandably, the aforementioned fusion weights can be obtained by normalizing the confidence set or by calibrating and mapping the confidence set, so as to make the fusion result more sensitive to high-confidence labels and reduce the perturbation intensity of low-confidence labels on the control prior parameters. In an optional implementation, when the overall confidence of the confidence set is lower than a preset threshold, the controller can perform derating processing on the control prior parameter set according to a preset conservative rule, so as to make the constraint boundary corresponding to the control prior parameter set more convergent or make the output change rate restriction more stringent, thereby suppressing the policy deviation caused by misjudgment of object attributes.

[0077] In one embodiment, the controller generates a control vector based on a set of prior control parameters and an observation sequence. The observation sequence is used to provide a state representation corresponding to the current control cycle, and the set of prior control parameters is used to provide a reference caliber, a weight caliber, and a constraint caliber to jointly determine the value of the control vector.

[0078] Specifically, the controller extracts the current state vector or state feature set from the observation sequence. The state feature set may include statistical features, trend features, or stage features obtained from time-series segments of sensor measurement data, and maintains its correlation with the description information of the unstructured object to interpret the selection basis of the control prior parameters within the control cycle. The controller then calculates the control error or deviation within the reference target or reference process range defined by the control prior parameter set, determines the control quantity allocation under the weighting caliber defined by the control prior parameter set, and performs amplitude limiting and rate of change constraint processing on the control quantity under the constraint caliber defined by the control prior parameter set, thereby generating a control vector. The control vector can be a one-dimensional or multi-dimensional set of control quantities and is used to characterize the target adjustment intensity or combination of adjustment actions on the controlled object.

[0079] In one embodiment, the controller converts the control vector into a drive command signal output to adapt to the interface constraints and safety constraints of the actuator. Specifically, the controller performs scaling, quantization, and encoding processing on the control vector according to the control interface type of the actuator to generate a drive command signal; the drive command signal may include at least one of pulse width modulation parameters, speed setpoint, switching pulse sequence, or command field in a communication message.

[0080] Before outputting the drive command signal, the controller can apply upper limit, lower limit and rate of change limits to the drive command signal based on the constraint parameters in the control prior parameter set. This is to avoid overshoot or frequent switching of the control output when there are fluctuations in the observed sequence, and to limit the deviation of the control strategy caused by misjudgment of object attributes to a controllable range, thereby improving the consistency and robustness of the control results.

[0081] In a high-configuration embodiment, in order to improve the physical interpretability of the observation sequence to the operating state and reduce the impact of the fluctuation of the original measurement quantity over time on subsequent inference and control, the controller performs physical state transformation on the time segment based on the sensor measurement data during the formation of the observation sequence to obtain the physical state sequence, and adds the physical state sequence as the physical representation of the time segment to the observation sequence, while maintaining the correlation between the physical state sequence and the unstructured object description information.

[0082] Specifically, after obtaining the sensor measurement data, the controller can first perform preprocessing on the sensor measurement data to form a measurement input that can be used for transformation. The preprocessing can include at least one of unit conversion, calibration correction, missing value imputation, outlier removal, and smoothing filtering. Subsequently, the controller performs physical state transformation on the measurement input at each sampling time or according to a preset window rule to obtain candidate physical state vectors and form a physical state sequence according to the sampling order.

[0083] The physical state sequence includes at least one physical state quantity, which is a state quantity or state index that can be calculated from sensor measurement data and has a defined physical semantics, such as a derived quantity or combination thereof used to characterize the environmental thermal and humidity state, energy state, or load state.

[0084] In one specific embodiment, physical state transformation is used to obtain a physical characterization of thermal and humidity states from environmentally relevant sensor measurement data. For example, the sensor measurement data includes at least one or a combination of temperature and humidity measurements. The controller maps the temperature and humidity measurements to derived state quantities related to thermal and humidity states based on preset physical relationships, lookup table rules, or a fixed calibration model, and writes these derived state quantities as physical state quantities into candidate physical state vectors. In an optional implementation, the mapping can be implemented using an enthalpy-humidity chart lookup table, an empirical fitting model, or a device factory calibration model. The lookup data, calibration parameters, or model parameters can be pre-stored in the controller's parameter area, enabling the controller to obtain a physical state sequence in real time at the edge according to the sampling time sequence. To balance computational overhead and timing stability, the controller can also perform segment-level aggregation on the physical state quantities according to the window range of the time sequence segments, based on the obtained sample-by-sample physical state vectors, to obtain segment-level physical state vectors. These segment-level physical state vectors then form a physical state sequence. Aggregation can be at least one of the following: mean, extreme values, quantiles, slope, or amplitude of change.

[0085] When adding the physical state sequence to the observation sequence, the controller maintains a consistent correspondence between the physical state sequence and the time segments of the sensor measurement data in terms of time index or window index, so that each observation unit contains not only the original or preprocessed time segment data representation, but also the physical state vector or its identifier that corresponds one-to-one with the time segment.

[0086] Furthermore, the controller maintains the association between the physical state sequence and the unstructured object description information unchanged. The association is inherited from the session identifier, control cycle identifier, timestamp alignment relationship, or time segment index established when the observation sequence is formed. For example, when a certain unstructured object description information is bound to a time segment, the physical state vector corresponding to the time segment is written into the same observation unit as the derived physical representation of the time segment and the unstructured object description information, or referenced in the observation sequence with the same association identifier, thereby avoiding mismatch between text semantics and physical representation within the control cycle.

[0087] Furthermore, physical state transformation can be achieved using at least one of the following alternative paths, without being limited to: generating energy state quantities or executing intensity state quantities based on relevant measurements such as power, current, rotational speed, or switching state; generating dynamic disturbance state quantities based on measurements such as vibration, noise, or position change; generating a comprehensive state index based on multi-channel measurements through a preset linear mapping, nonlinear mapping, or state observer; or combining multi-dimensional measurements of sensor data into a physical state vector according to preset normalization and scale alignment rules and outputting it as a physical state sequence.

[0088] By incorporating the physical state sequence into the observation sequence as a physical representation of time segments and maintaining consistency with the description information of unstructured objects, a more stable and interpretable state basis can be provided for subsequent model input data construction, semantic inference, and control vector generation.

[0089] Optional, see Figure 2 This application provides a method for generating model input data, including steps S201 to S203, wherein:

[0090] S201: Perform quantization encoding on the observed sequence to generate a time series labeled sequence;

[0091] S202: Obtain the running stage identifier determined by the preset running state machine;

[0092] S203: Construct the model input data based on the time series marker sequence, the running stage identifier, and the unstructured object description information.

[0093] Optionally, in order to enable the pre-trained semantic inference model to receive the temporal state information in the observation sequence and introduce the running phase context through a defined data interface, the controller performs quantization encoding on the observation sequence after forming the observation sequence to generate a time series label sequence, and obtains the running phase identifier determined by the preset running state machine, and then constructs the model input data based on the time series label sequence, the running phase identifier, and the unstructured object description information.

[0094] Specifically, the controller performs quantization encoding on time-series segments of sensor measurement data and / or the corresponding physical state sequences in the observation sequence to discretize continuous values ​​or multidimensional state vectors into a labeled sequence that can be received by the model interface. Quantization encoding can be performed at the sampling time-by-sampling granularity or the window-by-window granularity: at the sampling time-by-sampling granularity, the controller normalizes and bins the single-dimensional or multi-dimensional measurement at each sampling time to obtain the corresponding discrete label number; at the window-by-window granularity, the controller extracts statistical features, trend features, or segment-level feature vectors from the time-series segments under preset window length and step size constraints, and maps the feature vectors to discrete label numbers according to a preset codebook or binning threshold. The resulting time-series labeled sequence is arranged in sampling order or window order, so that each label can correspond one-to-one with its source sampling time or window position; when there are missing samples, abnormal samples, or incomplete windows, preset missing label, nearest neighbor completion, or window skipping rules can be used to handle the situation while maintaining the temporal consistency of the time-series labeled sequence.

[0095] In one embodiment, the controller further acquires the operation stage identifier determined by a preset operation state machine. The preset operation state machine is a stage determination logic pre-configured and fixed in the controller, which includes a finite set of operation stages and corresponding stage transition rules; the set of operation stages may include at least two of the initialization stage, operation stage, convergence stage and abnormal stage, or include multiple discrete stages used to distinguish different policy calibers.

[0096] The stage transition rule can be determined based on at least one of the following: runtime, threshold conditions or trend conditions of sensor measurement data or physical state sequences, event input, and drive command signals output by the controller. In an optional implementation, the controller updates the stage corresponding to the current time segment according to the stage transition rule in each control cycle and outputs a running stage identifier to identify the stage of the time segment. The running stage identifier can be an enumerated value, a discrete number, or an equivalent discrete identifier, and can be bound to the control cycle identifier or the time segment identifier to maintain a consistent correspondence between stage information and time segments when constructing subsequent model input data. In an optional implementation, the controller performs encoding processing on the running stage identifier when constructing model input data, converting it into a one-hot vector or an equivalent embedded vector representation, and concatenates or aligns it with the time series label sequence and the label sequence of the unstructured object description information, thereby enabling the pre-trained semantic inference model to achieve adaptive semantic inference using stage condition information.

[0097] Based on the above, the controller constructs model input data using time-series marker sequences, runtime phase identifiers, and unstructured object description information. The model input data is used to represent the joint information of object description semantics, discrete temporal state representation, and runtime phase context. Its construction can include inserting runtime phase identifiers as phase markers into preset positions in the model input data, or converting runtime phase identifiers into a phase marker sequence and combining it with the text marker sequence and time-series marker sequence to form the model input data.

[0098] For example, the controller can map the running phase identifier to a preset phase marker and concatenate them in the order of "phase marker - text marker sequence - separator marker - time series marker sequence" to form the model input data; or, the running phase identifier can be packaged as a global condition marker, text marker sequence, and time series marker sequence as input, and segment markers and position markers can be added to different modalities respectively, so that the model can distinguish between the text semantic dimension and the temporal state dimension and utilize the phase context information at the same time.

[0099] In addition, to meet the input length constraints of the pre-trained semantic inference model, the controller can truncate or pad the model input data according to preset priorities: when the length exceeds the limit, the text tag sequence associated with the current control cycle and the most recent time series tag are retained first; when the length is insufficient, preset padding tags are used to pad and an attention mask is generated to ensure the consistency of the inference interface and the determinism of the edge computing.

[0100] In this way, the model input data structurally includes time series label sequences organized in the sampling order, runtime phase identifiers representing the current control context, and unstructured object description information associated with time series segments. This enables the pre-trained semantic inference model to infer object semantics and output object semantic output under the runtime phase caliber, thereby providing a more stable and consistent input basis for determining the subsequent semantic label set and credibility set.

[0101] Optionally, in one embodiment, to enable the pre-trained semantic inference model to simultaneously perceive recent control action parameters when performing object semantic inference, the controller further introduces the actuator control instruction sequence from the previous control cycle during the generation of model input data, and encodes it into a control tag sequence before adding it to the model input data. The actuator control instruction sequence from the previous control cycle is a serialized record of the drive instruction signals output by the controller to the actuator in the previous control cycle. The drive instruction signals may include at least one of pulse width modulation parameters, speed setpoints, switching pulse sequences, or instruction fields in communication messages. The controller can cache the corresponding instruction parameters as a control instruction sequence in chronological order while outputting the drive instruction signal in each control cycle, so that subsequent control cycles can directly read the control instruction sequence from the previous control cycle as context information for constructing model input data.

[0102] In a preferred implementation, the controller performs normalization and quantization encoding on the control command sequence of the actuator to generate a control tag sequence. Specifically, the controller can determine a preset quantization caliber for each dimension of the command quantity in the control command sequence, including the value range, binning threshold, or codebook mapping table of that dimension of the command quantity; map the command quantity at each sampling time to discrete tag numbers according to the quantization caliber, and form a control tag sequence in chronological order. For multi-dimensional command quantities, the controller can generate corresponding discrete tag numbers for each dimension of the command quantity and combine them into a composite tag according to a preset splicing order, or arrange the multi-dimensional tags in an interwoven manner according to dimensions to form a control tag sequence, so that the control tag sequence can characterize the trend and intensity of the control action changes in the previous control cycle.

[0103] In one embodiment, the controller adds a control tag sequence to the model input data to form a joint input for semantic inference together with the text tag sequence and the temporal tag sequence. For example, the model input data includes text tag segments, temporal tag segments, and control tag segments. The controller can insert separator tags between the tag segments and set segment tags for different tag segments to indicate the tag source. The control tag segment corresponds to the control tag sequence of the previous control cycle, the temporal tag segment corresponds to the temporal tag sequence of the current control cycle, and the text tag segment corresponds to the text tag sequence of unstructured object description information associated with the temporal segment of the current control cycle.

[0104] Furthermore, to maintain the interpretable correspondence between different periodic information, the controller can attach boundary markers or periodic markers to the model input data to indicate the boundaries of the control period. It can also align the control marker sequence with the time sequence marker sequence based on preset alignment rules: for example, aligning them according to the number of sampling points within the control period, or resampling the control marker sequence over time to the same sequence length as the time sequence marker sequence. When the sequence length exceeds the maximum length allowed by the model interface, the controller can truncate earlier control markers or redundant time sequence markers according to preset priorities. When the sequence length is insufficient, it can be padded with padding markers and an attention mask can be generated so that the model input data meets the input interface constraints of the pre-trained semantic inference model.

[0105] By incorporating the sequence of actuator control commands from the previous control cycle into the model input data as a sequence of control tags, the pre-trained semantic inference model can simultaneously utilize three types of information—"object description semantics," "sensor temporal state," and "recent control actions"—for joint representation when inferring the semantic output of an object. This reduces the probability of misinterpreting state changes caused by control actions as changes in object attributes, even when sensor measurement data changes over time and control actions affect the measurement state. Consequently, it improves the stability of the semantic tag set and the credibility set, and provides a more consistent semantic basis for subsequent determination of control prior parameters and generation of driving commands.

[0106] Optionally, the pre-trained semantic inference model is an offline pre-trained semantic inference model with fixed base parameters. To adapt to the differences in the statistical characteristics of observed sequences and the semantic caliber of object descriptions at different operating stages, the pre-trained semantic inference model further includes multiple semantic adapter parameter sets. The semantic adapter parameter set is a set of switchable parameters attached to the base parameters of the pre-trained semantic inference model. It is used to modulate the intermediate representations of preset insertion points within the model without changing the main structure of the pre-trained semantic inference model, so that the object semantic output has a consistent interpretation caliber and stable output characteristics at different operating stages.

[0107] For example, a pre-trained semantic inference model may include a backbone network for generating joint representations and an output submodule for outputting object semantic outputs. A semantic adapter parameter set operates on one or more pre-defined insertion points in the backbone network, which may be located at least one of a text encoding path, a temporal encoding path, or a fusion path. The semantic adapter parameter set performs linear transformations, bottleneck transformations, gating modulation, or equivalent learnable modulations on the intermediate representations at the corresponding insertion points to adjust the model's fusion weights or representational emphasis on the text semantic dimension and the temporal state dimension, thereby forming an object semantic output that matches the runtime stage. The object semantic output may include a label score vector, a label probability vector, and / or confidence estimation components for a pre-defined semantic label space, enabling the subsequent determination of a set of semantic labels and a set of confidence levels corresponding to each semantic label set.

[0108] In one embodiment, multiple semantic adapter parameter groups are stored in the controller's model parameter area or policy library in the form of parameter packages or parameter tables. Each semantic adapter parameter group has a corresponding adapter identifier. The controller selects a target semantic adapter parameter group based on model input data, where the model input data includes a runtime phase identifier. Specifically, the controller can parse the runtime phase identifier from the model input data, determine the target adapter identifier according to a preset phase-adapter mapping relationship, load the target semantic adapter parameter group corresponding to the target adapter identifier, and keep the target semantic adapter parameter group fixed within the current control cycle and participate in forward inference calculation.

[0109] In one embodiment, the runtime phase identifier, in addition to being used for adapter selection, can also be encoded as a phase marker and added to the model input data, so that the pre-trained semantic inference model can use the runtime phase identifier to form a consistent phase semantic caliber when generating joint representations.

[0110] For example, the controller can encode the runtime phase identifier as a phase marker and insert it into a preset position in the text marker sequence and the time sequence marker sequence, or encode the runtime phase identifier as a phase embedding vector and align it with the joint representation of the model input data, so that the inference process inside the model and the adapter selection process outside the controller are consistent in terms of phase caliber.

[0111] In one embodiment, the controller generates object semantic output under the action of the target semantic adapter parameter set, and determines the semantic label set and the confidence set based on the object semantic output. When entering the next control cycle, the controller can reselect the target semantic adapter parameter set based on the new model input data, so that the semantic inference process can adaptively switch with the change of the running stage without introducing cross-cycle caliber drift, thereby providing a stable semantic basis for subsequently determining the control prior parameter set and generating the control vector based on the semantic label set and the confidence set.

[0112] Optional, see Figure 3 This application provides a method for obtaining a physical state sequence, including steps S301 to S302, wherein:

[0113] S301: Perform physical state transformation on the sensing measurement data at each sampling time to obtain candidate physical state vectors, and calculate physical consistency residuals based on the compatibility relationship between at least two types of redundant physical state quantities in the candidate physical state vectors.

[0114] S302: Based on the physical consistency residual, the candidate physical state vector is divided into steady-state points and disturbance points. After performing projection correction on the disturbance points under the preset physical feasible region constraint, they are reconstructed in segments with the steady-state points to obtain the physical state sequence.

[0115] Optionally, to reduce the impact of inconsistent state calibers introduced by multi-source sensor measurement data under conditions of noise, short-term disturbances, or measurement link anomalies on subsequent semantic inference and control vector generation, the controller can perform physical state transformation on the sensor measurement data when forming the observation sequence to obtain a physical state sequence. Further consistency checks and corrections are then performed on the physical state sequence to ensure that it meets preset physical feasible region constraints and has a stable consistency caliber. The input to the physical state transformation is multi-channel sensor measurement data aligned to the sampling time, and the output is a candidate physical state vector corresponding to each sampling time. The candidate physical state vector consists of at least one physical state quantity, and each dimension of the physical state quantity is used to characterize the physical state of the same controlled process. The physical state quantity can be a quantity directly measured by the sensor or a derived quantity derived from multiple measured quantities. Preferably, the candidate physical state vector contains at least two types of physical state quantities with redundant characterization relationships to cross-check the consistency of the state caliber without relying on a single measurement channel.

[0116] In one embodiment, the physical state transformation can be derived and combined from the sensor measurement data at each sampling time according to a preset transformation rule to generate a candidate physical state vector. The transformation rule can be determined based on device calibration parameters, sensor range parameters, or pre-configured state derivation parameters, and is stored in the controller parameter area or strategy library.

[0117] For example, the candidate physical state vector may include derived state quantities derived from measurements such as temperature, humidity, and air pressure, and / or operating state quantities derived from measurements such as motor speed, current, and load weight; it may also include two equivalent representations of the same state, such as one being a directly measured quantity and the other being an equivalent quantity derived from other measured quantities. Under normal physical conditions, the two should satisfy consistency constraints or boundary constraints, thus constituting "redundant physical state quantities".

[0118] It is understandable that "redundancy" does not refer to repeated sampling, but rather to the existence of a relationship between at least two types of physical state variables that can be derived from or constrain each other in a physical sense, making them usable for consistency verification.

[0119] In one embodiment, the controller calculates the physical consistency residual based on the compatibility relationship between at least two types of redundant physical state quantities in the candidate physical state vector. The compatibility relationship is used to characterize the consistency caliber that redundant physical state quantities should satisfy under physical constraints, and it can be at least one of the following: difference consistency relationship, relative consistency relationship, closed-loop consistency relationship, or feasible region boundary relationship. For example, when the candidate physical state vector contains both a type of derived quantity and its corresponding inverse quantity, the deviation between the two at the same sampling time can be used as the degree of violation of the compatibility relationship; when the candidate physical state vector contains both types of state quantities governed by the same physical constraint, the degree of exceeding the preset constraint boundary after their combination can be used as the degree of violation of the compatibility relationship. To make the residuals of different dimensions comparable, the controller can normalize the deviation, and the normalization factor can be determined by the range, allowable error band, or preset weight; and can perform time smoothing processing on the physical consistency residual, for example, taking the moving average or median of the residuals at several consecutive sampling times, to reduce the impact of single-point noise on the partitioning results.

[0120] In one embodiment, the controller divides candidate physical state vectors into steady-state points and disturbance points based on physical consistency residuals. A steady-state point represents a sampling moment where the candidate physical state vector satisfies the compatibility relationship and the residual is within a preset consistency condition range. A disturbance point represents a sampling moment where the candidate physical state vector does not satisfy the compatibility relationship or the residual exceeds the preset consistency condition range. The preset consistency condition can be determined by a fixed threshold, a stage threshold, or an adaptive threshold; wherein, the stage threshold can be selected based on the operating stage identifier, equipment specification parameters, or environmental parameters; the adaptive threshold can be determined based on the statistical distribution of the residuals, noise estimation results, or the residual level of historical stable periods.

[0121] In addition, to avoid frequent jitter near the threshold causing alternating switching between steady-state and disturbance points, the controller can introduce hysteresis rules, minimum duration rules, or minimum segment length rules. For example, the corresponding sampling point is only judged as a disturbance point when the residual exceeds the threshold for a preset number of consecutive times, or the steady-state segment is only confirmed when the number of consecutive steady-state points reaches a preset length.

[0122] In one embodiment, the controller performs projection correction on the disturbance point under a preset physical feasible region constraint. The preset physical feasible region is used to define the possible value range and variable range of the physical state vector, and it can be composed of at least one type of value boundary constraint and at least one type of consistency constraint. Value boundary constraints may include upper and lower limits of each physical state quantity, allowable rate of change limits, or restrictions related to the safety boundary of the actuator; consistency constraints may include allowable error bands corresponding to compatibility relationships, feasible intervals of combined constraints, or reference ranges defined by strategy library / calibration parameters. During projection correction, the controller can take "minimizing the deviation between the corrected physical state vector and the corresponding candidate physical state vector under the premise of satisfying the preset physical feasible region constraints" as the correction objective, and weight the deviations of each dimension based on preset weights; the weights can be determined by sensor reliability, range normalization weights, or fixed weights, so that more reliable or more sensitive state quantities have higher constraint strength in the correction. If a perturbation point does not have a feasible correction result under the current constraint set, the controller can perform a rollback process, such as marking the perturbation point as a missing point and filling it in by subsequent reconstruction, or selecting the feasible boundary point closest to the candidate physical state vector as the correction result, so as to ensure the continuity and availability of the physical state sequence output.

[0123] In one embodiment, after completing the projection correction of the disturbance points, the controller performs segmented reconstruction of the corrected disturbance points and steady-state points to obtain a physical state sequence. The segmented reconstruction combines the steady-state points and corrected disturbance points based on the sampling time sequence, ensuring that the output physical state sequence is consistent with the sampling order of the sensor measurement data in time, and maintains its association with the unstructured object description information in the observation sequence. During segmented reconstruction, the controller can determine the segment boundaries based on the adjacent switching positions of the steady-state points and disturbance points, and perform completion processing on missing points, isolated disturbance points, or short-term abnormal segments within the segment. The completion processing can employ at least one of the following: interpolation of nearby steady-state points, constraint-based smooth completion, or template completion based on historical stable segments, to prevent abnormal points from directly entering the subsequent semantic inference and control parameter determination links.

[0124] In this way, the resulting physical state sequence can maintain a consistent physical interpretation under perturbation conditions, thus providing a foundation for constructing more robust model input data based on the physical state sequence and generating more stable control outputs.

[0125] Optionally, in one embodiment, after completing the division of steady-state points and disturbance points and the projection correction of disturbance points, the controller performs segmented reconstruction processing on the corrected disturbance points and steady-state points to obtain a physically continuous physical state sequence that is convenient for subsequent evaluation and alignment. Specifically, firstly, the segment boundaries formed by the adjacent switching of steady-state points and disturbance points are identified in the sampling time sequence. The segment boundaries can be taken as the state points corresponding to the sampling time when the steady-state point marker and the disturbance point marker switch, or as the boundary pairs composed of the nearest steady-state points and disturbance points before and after the switching point. In an optional implementation, to avoid frequent switching caused by noise, the controller can merge or ignore short segments with a duration shorter than the preset minimum segment length, and can introduce a hysteresis criterion for the segment boundaries to keep the segment boundaries stable in adjacent control cycles.

[0126] After determining the segment boundaries, the controller constructs process features for boundary splicing and process consistency verification under the constraints of a preset reference process range. The preset reference process range defines the process caliber of the physical state vector under normal or desired operation. It can be composed of a set of reference trajectory bands, reference windows, or reference intervals pre-set by the controller and stored in the strategy library or parameter area. The preset reference process range can be selected according to the operating task, operating mode, environmental settings, or control prior parameter templates to ensure consistency of the reference caliber under different task conditions. Based on the physical state vector and the preset reference process range, the controller determines a first process feature and a second process feature. The first process feature characterizes the deviation driving force of the physical state vector relative to the reference process range, and the second process feature characterizes the degree of advancement of the physical state vector along the reference process range.

[0127] For example, the first process feature can be determined by the minimum distance from the physical state vector to the reference process range, the over-limit amplitude beyond the reference trajectory band, or their normalized results, and the deviation can be normalized by combining the reference band width or range to obtain a consistent numerical caliber; the second process feature can be determined by the matching position of the physical state vector within the reference process range, for example, matching the physical state vector to the nearest point on the reference trajectory corresponding to the reference process range, and using the arc length coordinate, stage percentage, or equivalent process coordinate corresponding to the nearest point as the degree of progress, so that the second process feature can reflect the process progress in the sampling time sequence.

[0128] In one embodiment, the controller applies boundary continuity constraints to the reconstructed physical state vector at the boundaries of each segment to suppress process rollback and state jumps caused by segment splicing. The boundary continuity constraints include constraints on the continuity and monotonicity of the second process feature, as well as constraints on the jump variables of the first process feature. Specifically, the controller selects adjacent sampling points or a boundary window of a preset length on both sides of the segment boundary, calculates the position difference of the second process feature before and after the boundary, and limits this position difference to prevent changes opposite to a preset direction of change. The preset direction of change is either an increasing or decreasing direction, and can be determined by the process coordinate convention of the reference process range. For example, the arc length coordinates of the reference process range can be defined as increasing from small to large, thus causing the second process feature to tend to change along the increasing direction as the sampling sequence progresses. The controller further calculates the jump variables of the first process features at the segment boundary. The jump variables can be taken as the difference of the first process features of adjacent points before and after the boundary, or as the difference of the mean of the first process features within the boundary window. The jump variables are constrained within a reference range defined by a preset reference process range. The reference range can be given by the trajectory band width, tolerance band, or preset threshold corresponding to the reference process range to limit abrupt changes in deviation intensity at the boundary splicing. If the boundary continuity constraint is not satisfied, the controller can backtrack and re-correct the projection correction results of disturbance points near the boundary, or perform local smoothing processing on the reconstruction results within the boundary window to satisfy the boundary continuity constraint.

[0129] In one embodiment, to uniformly represent the segmented reconstructed process position in the physical state space and reduce the time axis distortion introduced by sampling non-uniformity or disturbance correction, the controller performs arc length parameterization on the reconstruction results that satisfy the boundary continuity constraints within the physical state space composed of at least two types of physical state variables in the physical state vector, and resamples at equal arc length steps to obtain a physical state sequence. Specifically, the controller selects at least two types of physical state variables as the coordinate dimensions of the physical state space. The physical state variables can be any combination of temperature, humidity, pressure, airflow intensity, load characteristics, or derived state variables obtained through physical state transformation, and can perform dimensional normalization or weighted scaling on each dimension to obtain a consistent distance metric. The controller calculates the distance between adjacent reconstructed state points according to the sampling order and accumulates them to obtain arc length coordinates, so that the reconstruction results form discrete process curves in arc length coordinates.

[0130] Subsequently, the controller generates resampling positions on the arc length coordinate system with a preset equal arc length step size, and performs interpolation on the physical state vector corresponding to the arc length position to obtain the resampled state points. The interpolation method can be linear interpolation, piecewise spline interpolation, or an equivalent interpolation method. In an optional implementation, the controller can use the segment boundary points as forced alignment points for resampling, ensuring that the resampling results maintain a consistent arc length alignment at the segment boundaries. Thus, the controller obtains a physical state sequence arranged with equal arc length steps, providing a unified process coordinate basis for subsequent comparisons, alignment, and deviation calculations of process features, and reducing the instability of the reconstructed sequence caused by misjudgments of object attributes and disturbance corrections.

[0131] For example, segmented reconstruction can also employ the following alternative paths: segment boundary identification can be obtained by threshold hysteresis determination of physical consistency residuals, rather than solely relying on steady-state point / disturbance point marker switching; the deviation driving force of the first process feature can be determined by the projection of the multidimensional deviation vector onto a preset normal direction, or by the proportion of over-limit arc length; the advancement degree of the second process feature can be determined by the stage numbering of the reference process range division, or by the dynamic time warping matching position of the reference trajectory; the continuity criterion in the boundary continuity constraint can adopt the trend consistency criterion or the rate of change constraint criterion within the boundary window; the arc length distance measurement can adopt weighted distance or Mahalanobis distance to reflect the importance differences of different physical state quantities, and the equal arc length step size can be determined by a preset fixed value or adaptively by control period and computational resource constraints. All of the above alternative paths are used to achieve a unified expression of the segmented reconstruction results at the boundaries and in the process coordinates.

[0132] Optionally, in one embodiment, to suppress the continuous deviation of control prior parameters caused by misjudgment of object attributes and its amplification in subsequent control cycles, after outputting the control vector and completing the output of the drive command signal, the controller updates and collects sensor measurement data within a preset control cycle, and forms an updated physical state sequence for prior consistency verification based on the updated sensor measurement data. The preset control cycle can be consistent with the controller's control loop cycle or be a sub-window of it; for example, the controller continuously collects a set of sensor measurement data samples within a preset response window after outputting the drive command signal, and organizes them into updated time segments according to the sampling time order to reflect the immediate response state after the control output.

[0133] In one embodiment, the controller performs a physical state transformation on the updated sensor measurement data to obtain an updated physical state sequence. For example, the controller reuses the aforementioned physical state transformation process, mapping the updated time-series segments to a physical state vector sequence, and performs physical consistency residual calculation and steady-state / disturbance point partitioning on the updated physical state sequence when necessary, to reduce the impact of event disturbances or transient noise on consistency verification. Furthermore, projection correction and segmented reconstruction mechanisms can be reused to ensure that the updated physical state sequence maintains physical feasibility and temporal continuity under preset physical feasible region constraints.

[0134] In one embodiment, the controller determines a first process feature and a second process feature based on the updated physical state sequence. Within the physical state space, the controller reparameterizes the first and second process features to the arc length coordinates corresponding to the reference process range and calculates the deviation to determine the prior consistency index. The reference process range is the "reference process caliber" used by the controller to interpret the current control prior parameter set. It can be provided by a parameter template, rule base, or policy base of the control prior parameter set and can be defined in the physical state space as a reference trajectory band. The first process feature characterizes the degree of deviation of the updated physical state relative to the reference process range, and the second process feature characterizes the degree of advancement of the updated physical state along the reference process range. For example, the first process feature can be composed of the over-limit amplitude, over-limit direction, or normalized deviation of the updated physical state point from the boundary of the reference trajectory band, and the second process feature can be composed of the matching arc length position, stage coordinates, or progress amount monotonically corresponding to the arc length of the updated physical state point on the reference trajectory, thereby providing a "deviation intensity caliber" and a "process position caliber" respectively.

[0135] In one embodiment, when the controller performs reparameterization, it determines the matching arc length coordinates of each state point in the updated physical state sequence within the reference trajectory band corresponding to the reference process range, and maps the first process feature and the second process feature to a process feature trajectory that varies with the arc length coordinates. Specifically, the controller can determine the nearest neighbor matching point, the projected matching point, or the matching point that satisfies the physical feasible region constraint for each updated state point within the reference trajectory band, and take the arc length coordinates of the matching point on the reference trajectory as the corresponding process coordinates; subsequently, the first process feature and the second process feature are interpolated, resampled, or aligned according to the arc length coordinates to obtain a process feature trajectory consistent with the arc length coordinates of the reference process range, so that process features under different sampling rates or different update window lengths can be compared under a unified coordinate caliber.

[0136] In one embodiment, when calculating the deviation, the controller forms an operable cumulative metric based on the extent to which the process feature trajectory exceeds the reference process range. For example, at each arc length position, the controller determines whether the first and second process features fall within the allowable range defined by the reference process range. When any process feature exceeds the allowable range at the corresponding arc length position, the controller determines the extent of the exceedance at that arc length position and accumulates the exceedances in arc length order to obtain the deviation. The accumulation method can be direct accumulation of the exceedances, weighted accumulation according to a preset weight, or accumulation normalized to the arc length step size, thus allowing the deviation to reflect the overall degree of deviation of the updated physical state from the reference process range. The controller further determines a priori consistency index based on the deviation. The priori consistency index can be a normalized score of the deviation or a consistency judgment result obtained by comparing the deviation with a preset threshold, used to characterize the degree of consistency between the current control prior parameter set and the physical response after control output.

[0137] In one implementation, when the prior consistency index does not meet the preset consistency condition, the controller triggers prior adaptive fusion to generate a hybrid control prior parameter set, and generates subsequent control vectors based on the hybrid control prior parameter set. The preset consistency condition can consist of threshold conditions, segment conditions, or a combination of conditions related to the control cycle; for example, when the prior consistency index is lower than the preset consistency threshold, or the deviation exceeds the preset deviation threshold, or continuous exceedances occur within a preset arc length segment, it is determined that the preset consistency condition is not met and fusion is triggered. After triggering, the controller selects at least two semantic labels as candidate labels based on the confidence set, and determines the label weights corresponding to the candidate labels respectively; wherein, the label weights are jointly determined based on the confidence set, physical consistency residuals, and deviations, so that semantic inference confidence information and physical consistency information form a cooperative constraint in weight allocation.

[0138] In one implementation, when determining tag weights, the controller uses a set of confidence levels to form an initial weight caliber and introduces physical consistency residuals and deviations to correct the weights. For example, the controller can assign a baseline weight to candidate tags based on the set of confidence levels, and when the deviation or physical consistency residual increases, it can reduce the weight of the currently dominant tag and increase the weight of the candidate tags. Furthermore, it can apply preset lower and upper limits to the weights of each tag and perform normalization to ensure that the weights have a stable numerical caliber.

[0139] In one embodiment, when the controller generates a hybrid control prior parameter set, it retrieves or instantiates corresponding control prior parameter templates for each candidate semantic label according to a preset mapping rule, and performs weighted fusion or selective fusion based on weight thresholds for similar parameter items to obtain the hybrid control prior parameter set. The hybrid control prior parameter set may include parameter items for characterizing the range of the reference process, parameter items for characterizing the control calculation weight caliber, and constraint parameter items for characterizing output limiting and rate of change limits. After completing parameter fusion, the controller may apply preset consistency constraints and safety constraints to the hybrid control prior parameter set, such as pruning upper and lower limits of parameter item values, performing convergence processing on rate of change limits, or derating the gain caliber, to ensure that the hybrid control prior parameter set still has a conservative and stable control caliber under inconsistent conditions.

[0140] In one implementation, when generating control vectors in subsequent control cycles, the controller uses a hybrid control prior parameter set to replace the current control prior parameter set or as a modified prior to the current control prior parameter set, and reuses the aforementioned control vector generation process to form a control output that is more robust to object attribute uncertainties. For example, the controller writes the hybrid control prior parameter set into the currently effective parameter area and associates it with the control cycle identifier, so that subsequent control vector generation processes use the hybrid prior in terms of reference process range, weighting, and constraint caliber, thereby suppressing policy deviation and instability risks in the event of object attribute misjudgment or measurement timing fluctuations.

[0141] Optionally, in one embodiment, to enable comparable evaluation of the deviations between the first process feature and the second process feature under a unified process coordinate caliber, the controller pre-sets a reference process range. The reference process range characterizes the reference process evolution caliber under given control prior parameter set constraints, and can be implemented using a reference trajectory data structure sharing arc length coordinates. Specifically, the reference process range includes a first reference trajectory band set for the first process feature and a second reference trajectory band set for the second process feature. The first reference trajectory band is defined by the first reference trajectory and its corresponding first upper and lower bound trajectories, and the second reference trajectory band is defined by the second reference trajectory and its corresponding second upper and lower bound trajectories. The first and second reference trajectories share the arc length coordinates corresponding to the reference process range, enabling synchronous determination of the first and second process features at the same arc length position. Arc length coordinates can be constructed sequentially from discrete nodes of the reference trajectory and obtained by accumulating the distances between adjacent nodes; in an optional implementation, the arc length coordinates can be further normalized to a preset range to adapt to the scale differences of different devices or different operating stages.

[0142] The arc length coordinates of the reference process range are defined by a preset arc length grid or a unified index sequence. The first reference trajectory, the second reference trajectory, and their respective upper and lower bound trajectories are all stored in a one-to-one correspondence with the unified arc length grid, thereby realizing shared arc length coordinates.

[0143] In one embodiment, after obtaining the updated physical state sequence and determining the first process feature and the second process feature, the controller performs reparameterization on the first process feature and the second process feature to form a process feature trajectory. Specifically, the controller determines the arc length position sequence for the updated physical state sequence in the arc length coordinates of the reference process range, and maps the first process feature and the second process feature to the corresponding first process feature trajectory and second process feature trajectory, respectively. The determination of the arc length position sequence can adopt the following preferred example path: In the physical state space, each physical state vector in the updated physical state sequence is matched with the reference trajectory corresponding to the reference process range, and the reference trajectory node with the smallest distance to each physical state vector or that satisfies the preset matching criterion is determined. The arc length coordinates of the reference trajectory node are used as the corresponding arc length position, where the distance is the distance under the preset distance metric. When the reference trajectory node is a discrete point set, linear interpolation can be performed between adjacent nodes to obtain continuous arc length positions, and the arc length position sequence is kept monotonically changing along the sampling time sequence to avoid backtracking. Based on the arc length position sequence, the controller performs interpolation and resampling on the first process feature and the second process feature, so that the sampling points of the first process feature trajectory and the second process feature trajectory are aligned with the arc length grid of the reference process range, thereby realizing the alignment of process features under "shared arc length coordinates".

[0144] In one embodiment, when calculating the deviation, the controller performs a zone-fall determination on the first process characteristic trajectory and the first reference trajectory band, and on the second process characteristic trajectory and the second reference trajectory band, respectively. Specifically, at each arc length position, the controller determines whether the trajectory value of the first process characteristic trajectory falls within the allowable interval defined by the first lower bound trajectory and the first upper bound trajectory, and determines whether the trajectory value of the second process characteristic trajectory falls within the allowable interval defined by the second lower bound trajectory and the second upper bound trajectory; for any process characteristic trajectory exceeding the arc length position of the corresponding reference trajectory band, the controller determines the over-limit magnitude, which can be defined as the distance from the process characteristic trajectory value to its corresponding allowable interval boundary, or as a normalized over-limit amount relative to the allowable interval bandwidth. The controller further accumulates the over-limit amplitude at the over-limit arc length position according to the arc length to obtain the deviation degree; in an optional implementation, when the arc length grid step size is constant, the over-limit amplitude can be accumulated at equal step size; when the arc length grid step size is non-uniform, the over-limit amplitude can be multiplied by the difference between adjacent arc lengths and accumulated with weights so that the deviation degree reflects both the over-limit intensity and the process range of over-limit coverage; the deviation degree can be obtained by summing the over-limit accumulation terms of the first process feature and the over-limit accumulation terms of the second process feature, or by weighting the two according to preset weights.

[0145] In an optional implementation, the reference trajectories, upper bound trajectories, and lower bound trajectories of the first and second reference trajectory bands can be calibrated offline by the controller and stored in a policy library, and bound to the parameter templates of the control prior parameter set. When the control prior parameter set is switched or merged, the controller can synchronously switch the corresponding reference trajectory band parameters to maintain consistency in the deviation calculation. In addition to the preferred example path described above, the determination of the arc length position sequence can also be achieved by nearest neighbor lookup, projection matching, or matching based on preset physical feasible region constraints; the determination of the landing zone can also be achieved by piecewise linear interpolation with boundaries to adapt to the discretely stored upper and lower bound trajectories; the accumulation of deviation can also be achieved by approximating the area of ​​the over-limit arc length segment.

[0146] Example 2

[0147] This embodiment provides a practical application example of the adaptive control method based on multi-parameter sensing described in Embodiment 1.

[0148] In this embodiment, the intelligent clothes drying rack includes a controller, a sensor assembly, and an actuator assembly.

[0149] The sensor assembly is used to collect sensing and measurement data, including at least two of the following: ambient temperature, ambient relative humidity, duct temperature, fan current, heating component current, power supply voltage, and body vibration intensity.

[0150] The controller acquires sensing measurement data at a sampling period and outputs a driving instruction signal at a control period. For example, the sampling period is configured as 1 s, the control period is configured as 5 s, the time-segment window length is configured as 30 s, the window update step is configured as 5 s, and a control period identifier is assigned to each control period for data binding and traceability.

[0151] In one embodiment, the controller normalizes the unstructured object description information input by the user and forms a text token sequence.

[0152] The normalization process includes synonym merging and invalid symbol cleaning. For example, "Don't be too noisy" and "Quiet down" are merged into the same semantic segment, "Hurry up" and "As soon as possible" are merged into the same semantic segment, and duplicate punctuation marks or emojis irrelevant to the semantics are deleted. The construction of the text token sequence can adopt word segmentation or sub-word segmentation methods, and map the segmentation units to discrete token numbers; when the text length exceeds the preset upper limit, the recently input object description segment is preferentially retained and the earlier segment is truncated.

[0153] In one embodiment, the controller serializes and characterizes the time segments and forms a time series token sequence. For example, the multi-channel measurement values at each sampling moment are respectively quantized and encoded, and each channel measurement value is mapped to a discrete interval number according to a preset quantization scale, and the numbers of each channel at the same sampling moment are packed into time series tokens in a fixed channel order; or, fragment-level statistical features are extracted from the 30 s window as time series feature tokens, and the statistical features include at least one of mean, maximum value, minimum value, and change trend direction.

[0154] Among them, the quantization scale and the channel order are pre-fixed in the controller parameters and remain consistent in different control periods.

[0155] In one embodiment, the controller combines the text token sequence and the time series token sequence to form a model input.

[0156] The combination methods include: splicing method and alignment method.

[0157] When using the splicing method, a separator token is inserted between the text token sequence and the time series token sequence, and segment tokens and position tokens are respectively added to the two token sequences, so that the pre-trained semantic inference model can distinguish text semantics and time series states and maintain the consistency of the time series order.

[0158] When using alignment, the text tag sequence is bound to its associated temporal segment window, and window boundary markers are appended to the model input to indicate the temporal range corresponding to the text. When the combined input length exceeds the maximum length allowed by the model interface, the controller prioritizes retaining the temporal tag corresponding to the most recent control cycle and truncates redundant text segments according to a preset priority; when the input length is insufficient, it is padded with preset padding markers and an attention mask is generated to ensure the determinism of edge inference.

[0159] In one embodiment, the pre-trained semantic inference model is an inference model that is pre-trained offline and has its parameters fixed. Its structure includes at least a text encoding submodule, a temporal encoding submodule, a fusion submodule, and an output submodule.

[0160] The text encoding submodule generates text semantic representations, the temporal encoding submodule generates temporal state representations, and the fusion submodule fuses the two types of representations to obtain a joint representation. The fusion method includes concatenation fusion, gating fusion, or cross-attention fusion. The output submodule outputs object semantic output based on the joint representation, wherein the object semantic output includes at least the label score or label probability for a preset semantic label dictionary.

[0161] The preset semantic tag dictionary includes at least two of the following categories: silent preference, quick completion preference, energy consumption priority preference, high moisture absorption object, low moisture absorption object, large volume object, and small volume object.

[0162] The controller determines the semantic tag set based on the tag selection rules. The tag selection rules adopt the Top K rule and configure K to be 2 or 3, or adopt the threshold filtering rule and configure the threshold to be 0.55. When the combination of Top K and threshold filtering conditions is met, the tags with the highest scores and not lower than the threshold are retained first, and the tags are supplemented according to Top K when there are not enough tags.

[0163] The controller further generates a confidence set, which corresponds one-to-one with the semantic label set. The confidence set can be obtained from the label probabilities through calibration mapping. For example, the controller calls a pre-fixed calibration mapping table to perform piecewise linear mapping on the label probabilities, so that the confidence has consistent numerical semantics in different running stages; or, the confidence set can be directly given by the confidence estimation components output by the model and restricted to the range of 0 to 1.

[0164] In one implementation, the controller generates a set of control prior parameters according to a preset mapping rule.

[0165] The preset mapping rules are stored in the form of a mapping table or a parameter template dictionary. Each semantic tag corresponds to a set of control prior parameter templates. The control prior parameter templates include at least target range parameter items, output limit parameter items, and rate of change limit parameter items. For example, the control prior parameter templates include parameters such as the upper limit of the target fan speed, the lower limit of the target fan speed, the upper limit of the heating power, the upper limit of the heating power change rate, and the noise constraint weight.

[0166] The mapping table has the following field definitions: semantic tag identifier, parameter name, parameter default value, parameter allowed range, and parameter change rate limit.

[0167] As an example configuration, the upper limit of the target fan speed corresponding to the quiet preference can be configured to be 1100 rpm to 1300 rpm, the lower limit of the target fan speed corresponding to the quick completion preference can be configured to be 1300 rpm to 1600 rpm, the upper limit of the heating power corresponding to the high moisture absorption object can be configured to a normalized range of 0.70 to 0.90, and the upper limit of the heating power corresponding to the energy consumption priority preference can be configured to a normalized range of 0.45 to 0.60. All modes are uniformly configured with a heating power change rate limit of no more than 0.05 per control cycle.

[0168] The above values ​​are exemplary values, and specific values ​​can be configured according to product safety policies, device rated parameters and noise specifications.

[0169] When the set of semantic tags meets the single tag triggering condition, the controller selects the target semantic tag with the highest priority and a credibility not lower than the preset single tag threshold, and instantiates the corresponding control prior parameter template into the control prior parameter set.

[0170] When the semantic tag set triggers the multi-tag fusion condition, the controller selects at least two semantic tags and reads the corresponding parameter templates respectively, and performs fusion on the same type of parameter items to generate a hybrid control prior parameter set.

[0171] The fusion methods include credibility-weighted fusion and threshold selection fusion. When using credibility-weighted fusion, the credibility of each label is normalized to obtain the label weight, and similar parameter items are weighted to obtain the mixed parameters. When using threshold selection fusion, parameter items of high-credibility labels are selected first, and when the credibility of a high-credibility label is lower than a preset threshold, it degenerates into a conservative parameter item.

[0172] To avoid parameter drift caused by low-confidence inference, the controller can set an overall confidence threshold. When the maximum confidence of the semantic tag set is lower than this threshold, the output amplitude will be narrowed and the rate of change will be more strictly limited to form a conservative control caliber.

[0173] In one implementation, the controller generates a control vector based on a set of prior control parameters and an observation sequence, and outputs a drive command signal.

[0174] The control vector consists of two components: fan control quantity and heating control quantity, which are used to characterize the fan speed setpoint or PWM duty cycle setpoint and the heating power setpoint, respectively.

[0175] The controller extracts current state features from the observation sequence, including the window mean and upward trend of relative humidity, the window mean and downward trend of duct temperature, and the fluctuation range of fan current, and determines the control vector within the target range and constraint boundary defined by the control prior parameter set.

[0176] The controller converts control vectors into drive command signals. For example, it maps fan control quantities to PWM duty cycle or speed setpoint fields, and heating control quantities to power setpoint fields. Before outputting, it applies upper and lower limit amplitude and rate of change limits based on the control prior parameter set to avoid overshoot or frequent switching.

[0177] For example, the controller is an embedded processor or SoC with an ARM core (such as the Cortex M series or low-power Cortex A series), which has basic fixed-point arithmetic capabilities and can be equipped with an NPU or DSP to accelerate inference. The model parameters can be stored using 8-bit quantization or mixed precision to meet storage and latency requirements.

[0178] In one embodiment, the text encoding submodule is used to embed and context encode the text tag sequence obtained by text normalization. It may include a word embedding layer and a context encoding layer. The word embedding layer maps each text tag to a dense vector and superimposes positional encoding. The context encoding layer uses a lightweight Transformer encoder or a gated recurrent network to obtain the semantic representation of the text. For example, the number of layers in the context encoding layer can be one to four, the hidden dimension can be any preset value between 64 and 256, the number of attention heads can be one to eight, and the maximum length of the text can be configured to 16 to 128 tags according to the input habits of the end side. It can also truncate excessively long texts according to a preset priority or pad the texts with padding tags to meet the input interface constraints.

[0179] The temporal coding submodule is used to perform serialization modeling on temporal label sequences or temporal feature label sequences formed by temporal segments of sensor measurement data. It may include an input projection layer and a temporal modeling layer. The input projection layer is used to map multi-channel sampled values ​​or their quantized encoding to a unified dimension. The temporal modeling layer uses a combination of one-dimensional convolutional networks and gated recurrent units or a pure one-dimensional convolutional structure to extract trend and stage information. For example, the sampling frequency can be configured from 0.5Hz to 10Hz, the temporal window length can be configured from 10s to 120s and can be updated using a sliding window, the convolutional kernel size can be configured from 3 to 7, and the hidden dimension of the gated recurrent unit can be configured from 32 to 128, thereby obtaining window-level temporal state representation without depending on a specific sensor type.

[0180] The fusion submodule is used to fuse text semantic representation and temporal state representation into a joint representation. It can be implemented by gating fusion, that is, the gating network generates gating coefficients based on the two types of representations to adjust the proportion of semantic components and state components in the joint representation. Then, the joint representation is obtained through a feedforward network. The feedforward network can include 1 to 3 fully connected layers and can add a normalization layer to improve numerical stability. In the terminal scenario, the dimension of the joint representation can be configured to be 64 to 256.

[0181] The output submodule is used to generate object semantic output based on joint representation, including a label score generation head and a confidence estimation head. The label score generation head maps the joint representation to a label score vector with the same dimension as the preset semantic label dictionary and obtains a label probability vector. The semantic label dictionary can be set to several label items according to the common object attributes of clothes dryers and user intent, for example, it can be 8 to 24 label items and include at least a part of the object attribute class labels and preference intent class labels.

[0182] The controller can determine the semantic label set by threshold filtering or Top K rule, and can take Top K as 2 or 3, and set the label probability threshold to any preset value between 0.50 and 0.80, so as to retain multiple candidate labels when the object attributes are uncertain; the confidence estimation head can directly output the confidence estimation component or the uncertainty estimation component. The controller can calibrate the label probability with offline fixed calibration parameters and use it as the confidence set, or convert the uncertainty estimation component into a confidence set according to the preset mapping calibrator, and limit the confidence to the range of 0 to 1 so that the subsequent control mapping module can use it uniformly.

[0183] Furthermore, the controller can encapsulate the semantic tag set, the confidence set, and the control cycle identifier into a semantic inference result structure and output it to the control prior parameter determination module, so that the interface relationship between the semantic inference output and the preset mapping rules is clear, traceable, and can maintain a consistent interpretation under different control cycles.

[0184] Optionally, the text encoding submodule can be implemented by a convolutional text encoder or a bidirectional recurrent network; the temporal encoding submodule can be implemented by a pure Transformer temporal encoder or a statistical feature encoder; the fusion submodule can be implemented by concatenation fusion or cross-attention fusion; and the output submodule can only output the label probability and the controller determines the credibility based on the concentration of the probability distribution. Moreover, the above exemplary parameters and structures are only used to illustrate the feasible path. Those skilled in the art can make equivalent substitutions and engineering adjustments to the number of layers, dimensions, window length, truncation and padding strategies and the number of labels without departing from the technical concept of this application.

[0185] Example 3

[0186] Based on the same inventive concept, this embodiment provides an adaptive control system based on multi-parameter sensing, corresponding to an adaptive control method based on multi-parameter sensing. Since the principle of the system in this embodiment for solving the problem is similar to that of the adaptive control method based on multi-parameter sensing described in Embodiment 1, the implementation of the system in this embodiment can refer to the implementation of the method, and the repeated parts will not be described again.

[0187] Reference Figure 4 This embodiment provides an adaptive control system based on multi-parameter sensing, including:

[0188] The acquisition module 10 is used to acquire sensor measurement data and unstructured object description information provided by the user to form an observation sequence;

[0189] Processing module 20 is used to generate model input data based on the observation sequence, input the model input data into a pre-trained semantic inference model to obtain object semantic output, and determine a set of semantic labels and a set of credibility that corresponds one-to-one with the set of semantic labels based on the object semantic output.

[0190] The output module 30 is used to determine a set of control prior parameters based on the set of semantic tags and the set of confidence according to a preset mapping rule, generate a control vector based on the set of control prior parameters and the observation sequence, and convert the control vector into a driving command signal for output.

[0191] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method of adaptive control based on multi-parameter perception, characterized in that, The method comprises: acquiring sensor measurement data and user-provided unstructured object description information to form an observation sequence; based on the observation sequence, generating model input data, inputting the model input data into a pre-trained semantic inference model to obtain object semantic output, determining a semantic tag set and a confidence set corresponding to the semantic tag set according to the object semantic output; based on the semantic tag set and the confidence set, determining a control prior parameter set according to a preset mapping rule, generating a control vector based on the control prior parameter set and the observation sequence, and converting the control vector into a driving instruction signal output.

2. The method of claim 1, wherein, The observation sequence includes a time sequence segment of the sensor measurement data and unstructured object description information associated with the time sequence segment. Alternatively, the observation sequence includes a time sequence segment of the sensor measurement data, unstructured object description information associated with the time sequence segment, and a physical state sequence. The physical state sequence is a physical characterization of the time sequence segment, and at least one physical state quantity is obtained by performing a physical state transformation on the time sequence segment.

3. The method of claim 1, wherein, The generation of the model input data includes: performing quantization encoding on the observation sequence to generate a time sequence label sequence; acquiring a running phase identifier determined by a preset running state machine; based on the time sequence label sequence, the running phase identifier, and the unstructured object description information, constructing the model input data.

4. The method of claim 1, wherein, The generation of the model input data also includes: acquiring an actuator control instruction sequence of the previous control cycle, encoding the actuator control instruction sequence to generate a control label sequence, and adding the control label sequence to the model input data.

5. The method of claim 1, wherein, The pre-trained semantic inference model includes a plurality of semantic adapter parameter groups, a target semantic adapter parameter group is selected based on the model input data, and the object semantic output is generated under the action of the target semantic adapter parameter group.

6. The method of claim 2, wherein, The physical state transformation includes: performing a physical state transformation on the sensor measurement data at each sampling time to obtain a candidate physical state vector, and calculating a physical consistency residual based on the compatibility relationship between at least two types of redundant physical state quantities in the candidate physical state vector; according to the physical consistency residual, dividing the candidate physical state vector into steady points and disturbance points, and performing projection correction on the disturbance points under the constraint of a preset physical feasible region, and then segmenting and reconstructing with the steady points to obtain the physical state sequence.

7. The method of adaptive control based on multi-parameter perception according to claim 6, characterized in that, The segmentation and reconstruction includes: determining a segment boundary formed by the adjacent switching of the steady points and the disturbance points, determining a first process feature representing the deviation driving force of the physical state vector relative to a reference process range and a second process feature representing the advancement degree of the physical state vector along the reference process range based on the physical state vector and the preset reference process range; applying a boundary continuity constraint on the reconstructed physical state vector at the segment boundary, the boundary continuity constraint comprising: the second process feature being continuous at the segment boundary and having no change in the sampling time sequence in a direction opposite to a preset change direction, the preset change direction being one of an increasing direction or a decreasing direction, and constraining a jump of the first process feature at the segment boundary to be within a reference range defined by the preset reference process range; performing arc length parameterization on the reconstructed result satisfying the boundary continuity constraint in a physical state space constituted by at least two types of physical state quantities in the physical state vector, and resampling in equal arc length steps to obtain the physical state sequence.

8. The method of adaptive control based on multi-parameter perception according to claim 7, characterized in that, Further comprising: updating the sensing measurement data within a preset control period after outputting the control vector, and performing the physical state transformation on the updated sensing measurement data to obtain an updated physical state sequence; determining the first process feature and the second process feature based on the updated physical state sequence, and after re-parameterizing the first process feature and the second process feature to arc length coordinates corresponding to the reference process range in the physical state space, calculating a deviation degree of the first process feature and the second process feature relative to the reference process range, determining a prior consistency index based on the deviation degree; in response to the prior consistency index not satisfying a preset consistency condition, selecting at least two semantic labels based on the credibility set and determining label weights respectively, performing weighted fusion on control prior parameters corresponding to the at least two semantic labels according to the preset mapping rule to generate a mixed control prior parameter set, and generating a subsequent control vector based on the mixed control prior parameter set; wherein the label weights are determined based on the credibility set, the physical consistency residual, and the deviation degree.

9. The method of adaptive control based on multi-parameter perception according to claim 8, characterized in that, The reference process range includes a first reference trajectory band set for the first process feature and a second reference trajectory band set for the second process feature; the first reference trajectory and the second reference trajectory share the arc length coordinates corresponding to the reference process range; The first reference trajectory band is defined by a first upper boundary trajectory and a first lower boundary trajectory in the arc length coordinates of the first reference trajectory, and the second reference trajectory band is defined by a second upper boundary trajectory and a second lower boundary trajectory in the arc length coordinates of the second reference trajectory; calculating the deviation degree comprises: taking the first process feature re-parameterized to the arc length coordinates corresponding to the first reference trajectory band as a first process feature trajectory, and determining whether the first process feature trajectory falls within the first reference trajectory band at each arc length position; taking the second process feature re-parameterized to the arc length coordinates corresponding to the second reference trajectory band as a second process feature trajectory, and determining whether the second process feature trajectory falls within the second reference trajectory band at each arc length position; for any arc length position where a process feature trajectory exceeds the corresponding reference trajectory band, determining an out-of-limit amplitude and accumulating the deviation degree in arc length.

10. A multi-parameter perception based adaptive control system for implementing the multi-parameter perception based adaptive control method of any one of claims 1-9, characterized in that, Comprising: An acquisition module is configured to acquire sensor measurement data and unstructured object description information provided by a user to form an observation sequence; A processing module is configured to generate model input data based on the observation sequence, input the model input data into a pre-trained semantic inference model to obtain object semantic output, determine a semantic label set and a confidence set corresponding to the semantic label set according to the object semantic output; An output module is configured to determine a control prior parameter set according to a preset mapping rule based on the semantic label set and the confidence set, generate a control vector based on the control prior parameter set and the observation sequence, and convert the control vector into a driving instruction signal for output.

Images