A method for evaluating the operation of IoT devices based on digital twins

Abstract

This invention discloses a method for evaluating the operation of IoT devices based on digital twins, belonging to the field of digital twin device operation evaluation technology. The method includes the following steps: organizing the device's digital twin to form an operation migration chain; dividing the device's measured feedback content into operation support pieces to form migration intervals; connecting the migration intervals to an improved neural controlled differential equation model; arranging the migration intervals into adjacent integration intervals and advancing them to an exit hidden state; truncating the original interface of directly writing the initial value of the next integration interval from the exit hidden state, and pushing the exit change into the initial value receiving base; encapsulating closed changes into closed initial value components, and labeling unclosed changes as stagnant initial value components; continuing to advance the closed initial value components, and writing the stagnant initial value components back to the digital twin to form the operation evaluation result. This invention achieves closed-loop evaluation of the operational authenticity of IoT devices, migration boundary tracing, and collaborative anomaly write-back.

Description

Technical Field

[0001] This invention relates to the field of digital twin device operation evaluation technology, and in particular to a method for evaluating the operation of Internet of Things (IoT) devices based on digital twins. Background Technology

[0002] Existing methods for evaluating the operation of IoT devices typically rely on device online status, alarm records, sensor values, and feedback information for judgment. Some systems introduce digital twins to map and monitor device status. These methods can reflect whether the device is online, whether the status values ​​are abnormal, and whether feedback has been returned. However, the evaluation objects are mostly concentrated on the status of a single device or at a single moment, making it difficult to fully reflect the state transition relationships during device operation.

[0003] With the application of digital twin technology in IoT operation and maintenance, some systems are able to synchronize the status of physical devices to a virtual model and predict or evaluate the device's operating status. However, existing methods typically process device feedback, physical status changes, twin state migration, and adjacent device handover results separately, lacking continuous verification of the operational migration chain. This can easily lead to misjudging situations such as a completed feedback but unchanged physical status, or a changed physical status but not synchronized twin state migration, as normal operation.

[0004] Furthermore, existing time series models typically advance hidden states continuously along the time sequence, which can easily absorb disconnections, lags, and pseudo-completions at migration boundaries into continuous state changes. This results in smoothing out operational breakpoints and makes it difficult to accurately identify actual operational disconnections, twin migration stagnation, and collaborative acceptance anomalies.

[0005] Therefore, how to provide a method for evaluating the operation of IoT devices based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for evaluating the operation of IoT devices based on digital twins. This invention employs a combination of operational migration chains and an improved neural controlled differential equation model. By organizing the expected state migrations and collaborative acceptance boundaries in the digital twin of the IoT device, an operational migration chain capable of supporting real-world operational verification is constructed. The measured feedback content from the device is segmented into operational support pieces and pressed onto the corresponding migration boundaries, forming migration intervals carrying real-world operational support. Based on this, an initial value update layer is embedded at the intersection of adjacent integration intervals of the differential equation solver. The original interface for directly writing the initial value of the next integration interval from the hidden state of the previous integration interval is truncated. Real-world closed-loop changes and migration stagnation changes are distinguished through the initial value acceptance basis, closed initial value components, and stagnant initial value components. This enables accurate identification of IoT device operational disconnection, twin migration stagnation, and collaborative acceptance anomalies. It has the advantages of high evaluation result accuracy, traceable migration boundaries, and applicability to multi-device collaborative operation scenarios.

[0007] An IoT device operation evaluation method based on digital twin according to an embodiment of the present invention includes the following steps: The operation migration of digital twins of IoT devices is organized to align the expected state migration with the collaborative acceptance boundary, forming an operation migration chain; The actual measured data transmitted back from the equipment is divided into operation support pieces along the operation migration chain, and the migration boundaries are pressed together to form a migration range carrying real operation support. The migration interval access improves the neural controlled differential equation model by embedding an interval initial value update layer at the intersection of adjacent integration intervals of the differential equation solver. The migration intervals are arranged as adjacent integration intervals along the running migration chain. Each integration interval takes over the hidden entry state and advances along the interval state change to form the hidden exit state. At the junction of adjacent integration intervals, the original interface of the next integration interval initial value is truncated after the exit hidden state is directly written. The initial value is reset with the entry hidden state as the base, and the change of the exit hidden state relative to the entry hidden state is pushed into the initial value base to form the initial value update object. For changes in the initial value update object that are closed with the running support piece, the base is retained; for changes that are not closed with the running support piece, the boundary is rolled back, forming closed initial value components and retained initial value components. After the initial value component is connected to the integration interval, the advancing state and the stagnant initial value component are merged into a hidden state chain. The boundary assignment of the stagnant initial value component is marked and written back to the digital twin to form the operation evaluation result of the Internet of Things device.

[0008] Optionally, forming the runtime migration chain includes: Extract device state nodes with state transition tags from the digital twin of IoT devices, and pair device state nodes of the same device in adjacent operation phases according to the order of occurrence. The verification criteria are whether the states before and after the migration appear in pairs. The state migration pairs that appear in pairs are closed as the expected state migration, and the positions where the state gaps are located are retained as unclosed migration boundaries. In the digital twin, the task succession relationship between adjacent devices is checked, and the task succession relationship that the completed state of the previous device can continue the state of the next device is connected between the corresponding expected state transitions to form a collaborative succession boundary. The expected state transition of the unconnected task acceptance relationship is retained as the single device migration boundary, and the task acceptance relationship that cannot be continued by the state of the previous device is marked as the disconnection acceptance boundary. The desired state transitions are arranged in the order in which they occur and connected between adjacent desired state transitions according to their boundary categories, forming a running transition chain.

[0009] Optionally, the formation of the migration range carrying actual operational support includes: Extract the device affiliation and transmission sequence from the actual measured data transmitted back from the device. Arrange the state changes of the same device in adjacent transmission times according to the order of occurrence and attach them to the same device migration segment in the running migration chain. Cut the measured feedback content of the equipment in the same equipment migration segment along the migration boundary in the running migration chain, leave the feedback state before the migration boundary on the front side of the migration, leave the feedback state after the migration boundary on the back side of the migration, and retain the order of occurrence before and after the migration boundary to form the boundary feedback content. Verify the state changes in the boundary feedback content with the expected states on both sides of the migration boundary. When the state changes can be cut into running support pieces from the state before migration to the state after migration. When a state change cannot be continued from the pre-migration state to the post-migration state, the boundary backhaul content is retained at the migration boundary and marked as a support gap. The running support piece is pressed onto the corresponding migration boundary, and the support gap is left at the corresponding migration boundary, thus forming a migration range carrying the actual running support.

[0010] Optionally, the improved neural controlled differential equation model includes a control path, an initial network, a vector field network, a differential equation solver, an interval initial value update layer, and a readout layer, with the interval initial value update layer embedded at the intersection of adjacent integration intervals of the differential equation solver.

[0011] Optionally, the formation of the outlet concealment state includes: The migration intervals carrying real operational support are arranged into adjacent integration intervals along the occurrence order of the operational migration chain, and the migration boundaries between adjacent integration intervals are aligned as integration handover boundaries. The control path intercepts the state change sequence within each integral interval along the integral junction boundary, compresses the continuous change from the state before the migration to the state after the migration into the control path increment, and forms a continuous control segment with the same sequence according to the arrangement order of adjacent integral intervals. The initial network receives continuous control segments of the first integration interval and compresses the pre-transition state of the first integration interval into an entry hidden state. The vector field network reads the entry hidden state and continuous control segment within the differential equation solver, and applies the control path increment in the continuous control segment to the hidden dimension corresponding to the entry hidden state in turn, forming an interval advance quantity in the same order as the control path increment. The differential equation solver accumulates the interval advance quantity along the occurrence order of the continuous control segment. The hidden state at the entrance is accumulated through the interval advancement to form the corresponding hidden state at the exit, and maintains a corresponding relationship with the integral intersection boundary.

[0012] Optionally, the initial value update object includes: The interval initial value update layer intercepts the direct write interface between the previous integral interval exit hidden state and the next integral interval initial value entry at the integral handover boundary, closes the transmission end of the exit hidden state into the initial value entry as is, and retains the integral correspondence between the previous integral interval entry hidden state and the continuous control segment. The initial value basis is established based on the hidden state of the previous integration interval entry. The initial value basis is expanded sequentially along the hidden dimension of the entry hidden state. The initial value entry of the next integration interval is then connected to the initial value basis. The control path increments along the continuous control segment sequentially call the cumulative results formed by the vector field network in the differential equation solver. The hidden changes generated by the hidden state at the exit of the previous integral interval relative to the hidden state at the entrance of the previous integral interval are decomposed back to the source of the control path increment according to the hidden dimension corresponding to the control path increment, forming an interval change sequence consistent with the control path increment sequence. Align the interval change sequence with the same hidden dimension of the initial value receiving base in the order of control path increment, check the dimension correspondence between the interval change and the state before migration in the initial value receiving base item by item, and write the interval change sequence into the initial value receiving base. The source of the initial value of the next integration interval is changed from direct writing of the exit hidden state to receiving and superimposing the interval change in the entry hidden state. The initial value basis after writing the interval change sequence is dimension-merged. The merged initial value basis replaces the hidden state of the previous integration interval exit and becomes the source of the initial value for the next integration interval. The direct write interface from the hidden state of the previous integration interval exit to the initial value entry of the next integration interval is closed, forming an initial value update object carrying the initial value basis and the interval change sequence.

[0013] Optionally, the formation of the closed initial value component and the retention initial value component includes: The initial value inheriting base and interval change sequence in the initial value update object are inherited, and the interval change sequence is aligned with the corresponding migration boundary running support piece item by item along the control path incremental occurrence order; Using the ability of the running support piece to continue from the state before migration to the state after migration as the closure criterion, the interval changes where the closure criterion is valid are retained in the same-order hidden dimension of the initial value basis. Using the non-formed state of the running support piece as the basis for non-closure, the interval change where the non-closure basis is established is stripped from the corresponding hidden dimension of the initial value receiving base and returned to the migration boundary. The interval changes retained within the initial value basis are merged sequentially along the hidden dimension, and the merged initial value basis is closed into closed initial value components. The interval changes of the retreat migration boundary are arranged sequentially along the control path increments and maintain their correspondence with the migration boundary, forming the stagnant initial value component.

[0014] Optionally, the formation of IoT device operation evaluation results includes: The closed initial value component is connected to the hidden state at the entrance of the next integral interval, and the process continues along the continuous control segment of the next integral interval to form a chain of continuous hidden states. The stagnant initial value component is attached to the corresponding transition boundary in the continuing hidden state chain, while maintaining the order of occurrence between the stagnant initial value component and the corresponding interval change; The readout layer reads the transition boundary with the lingering initial value component along the continuous hidden state chain and verifies the state continuity relationship between the running support piece at the transition boundary and the desired state transition. When the state continuity relationship is established, the initial value component of the lingering state is marked as the twin migration lingering state; when the state continuity relationship is not established, the initial value component of the lingering state is marked as the actual operation disconnection state. The twin migration retention type and the real operation disconnect type are written back to the operation migration chain in the digital twin to form the operation evaluation results of IoT devices.

[0015] The beneficial effects of this invention are: This invention, by combining the operational migration of digital twins with an improved neural controlled differential equation model, enables continuous and traceable authenticity assessment of the operation process of IoT devices. By aligning the expected state transition with the collaborative acceptance boundary into an operational migration chain, and by segmenting the actual measured feedback content from the device into operational support pieces and pressing them onto the migration boundary, this invention unifies device feedback, physical state changes, twin state transitions, and adjacent device acceptance relationships into a single migration chain for verification. This avoids relying solely on online status, alarm records, or single feedback to judge device operation, thereby improving the completeness and accuracy of operational assessment.

[0016] Furthermore, this invention embeds an interval initial value update layer into the improved neural controlled differential equation model, truncating the original interface of directly writing the initial value of the next integration interval from the hidden state at the exit of the previous integration interval. This separates and processes the actual closed and unclosed changes during the continuous advancement of the hidden state. Compared with the way ordinary time series models continuously absorb state changes, this invention can retain disconnection, hysteresis, and pseudo-completion information at the migration boundary, reducing the problem of run-through breakpoints being smoothed by hidden states and improving the ability to identify actual run-through disconnections and twin migration stagnation.

[0017] This invention can also assign the stagnant initial value components to the corresponding migration boundaries and write them back to the digital twin, enabling IoT device operation evaluation results to not only provide anomaly conclusions but also locate the migration boundaries and transition points where the anomalies occurred. Therefore, this invention is applicable to multi-device collaborative operation scenarios, improving the authenticity of device operation evaluations, boundary traceability, and collaborative anomaly identification capabilities, providing a reliable basis for IoT system operation, scheduling, and state correction. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an IoT device operation evaluation method based on digital twin proposed in this invention; Figure 2 This is a schematic diagram of the structure of an improved neural controlled differential equation model for an IoT device operation evaluation method based on digital twins proposed in this invention. Figure 3 This is a schematic diagram of the structure of the interval initial value update layer rewriting the direct write initial value interface of the IoT device operation evaluation method based on digital twin proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A method for evaluating the operation of IoT devices based on digital twins includes the following steps: The operation migration of digital twins of IoT devices is organized to align the expected state migration with the collaborative acceptance boundary, forming an operation migration chain; The actual measured data transmitted back from the equipment is divided into operation support pieces along the operation migration chain, and the migration boundaries are pressed together to form a migration range carrying real operation support. The migration interval access improves the neural controlled differential equation model by embedding an interval initial value update layer at the intersection of adjacent integration intervals of the differential equation solver. The migration intervals are arranged as adjacent integration intervals along the running migration chain. Each integration interval takes over the hidden entry state and advances along the interval state change to form the hidden exit state. At the junction of adjacent integration intervals, the original interface of the next integration interval initial value is truncated after the exit hidden state is directly written. The initial value is reset with the entry hidden state as the base, and the change of the exit hidden state relative to the entry hidden state is pushed into the initial value base to form the initial value update object. For changes in the initial value update object that are closed with the running support piece, the base is retained; for changes that are not closed with the running support piece, the boundary is rolled back, forming closed initial value components and retained initial value components. After the initial value component is connected to the integration interval, the advancing state and the stagnant initial value component are merged into a hidden state chain. The boundary assignment of the stagnant initial value component is marked and written back to the digital twin to form the operation evaluation result of the Internet of Things device.

[0021] In this embodiment, the results of the IoT device operation evaluation include: Extract device state nodes with state transition tags from the digital twin of IoT devices, and pair device state nodes of the same device in adjacent operating phases according to the order of occurrence. Specifically, pairing device state nodes of the same device in adjacent operating phases according to the order of occurrence involves: The system reads the state records registered within the digital twin. These records include device identity, operational phase, state content, and occurrence time. Device identity is the device name or device code assigned to the physical IoT device by the digital twin. The operational phase is the device's state stage within a task. State content includes one of the following operational states: on, off, occupied, released, accessed, exited, started, or stopped. When the operational phase changes in the state record, the record is extracted as a device state node with a state transition marker.

[0022] Device status nodes with consistent device identities are grouped into the same device node set and sorted by occurrence time in ascending order. When occurrence times are the same, the order is determined by the write time of the digital twin. After sorting, adjacent device status nodes are paired, with the preceding node serving as the pre-migration node and the following node as the post-migration node, forming a state transition pair. Status records lacking device identity are categorized as unassigned records, and status records missing during operation are categorized as pending verification sets. Unassigned records and pending verification sets do not participate in the pairing of state transition pairs.

[0023] Using the presence of paired states before and after the transition as the verification criterion, state transitions where states appear in pairs are considered closed as expected state transitions, and the locations of state gaps are retained as unclosed transition boundaries. Specifically, the process of identifying closed state transition pairs as expected state transitions and retaining the locations of state gaps as unclosed transition boundaries is as follows: Read the pre-migration and post-migration nodes in the state transition pair, and verify the device identity, operational stage, state content, and state transition marker of the two nodes. If the device identity matches, both nodes contain state content, and the state transition marker points from the pre-migration node to the post-migration node, then the pre-migration and post-migration states are considered to exist in pairs. The state transition marker pointing from the pre-migration node to the post-migration node means that the state transition relationship recorded within the digital twin can convert the state content of the pre-migration node into the state content of the post-migration node.

[0024] When states occur in pairs, the state content of the node before migration is written into the pre-migration state of the expected state migration, and the state content of the node after migration is written into the post-migration state of the expected state migration. The order in which the two nodes occur is retained as the running order of the expected state migration. When a node before migration is missing, a node after migration is missing, the device identity is inconsistent, the state migration marker direction is broken, or the state content is missing, the position of the missing or broken node is recorded as a state gap. The state gap is fitted to the boundary position of the corresponding state migration pair to form an unclosed migration boundary.

[0025] In the digital twin, the task succession relationship between adjacent devices is verified. The task succession relationship where the completion state of the previous device can continue into the entry state of the next device is integrated into the corresponding expected state transitions, forming a collaborative succession boundary. Specifically, integrating the task succession relationship where the completion state of the previous device can continue into the entry state of the next device into the corresponding expected state transitions involves: Read the task succession relationships registered in the digital twin. The task succession relationship includes the previous device, the next device, and the task direction. The previous device is the device that completes the current action in the task chain, the next device is the device that takes over the result of the current action, and the task direction is the direction from the previous device's completed state to the next device's entering state.

[0026] Extract the expected state transitions for the preceding and following devices. If the post-transition state of the preceding device is a completed state, and the pre-transition state of the following device is an entering state, and the post-transition state of the preceding device occurs before the pre-transition state of the following device, then the task acceptance relationship can be determined to be continuous. This continuous task acceptance relationship is then placed between the expected state transitions of the preceding and following devices, forming a collaborative acceptance boundary. If a task acceptance relationship lacks a preceding device, a following device, or a task direction, the task acceptance relationship is categorized into the missing acceptance set, and no collaborative acceptance boundary is formed.

[0027] The expected state transition of a task acceptance relationship that has not been connected is retained as a single-device migration boundary. The task acceptance relationship where the previous device's completion state cannot be continued into the state of the next device is marked as a disconnection boundary. Specifically, the expected state transition of a task acceptance relationship that has not been connected is retained as a single-device migration boundary, and the task acceptance relationship where the previous device's completion state cannot be continued into the state of the next device is marked as a disconnection boundary. Read all expected state transitions and verify whether they are connected to task acceptance relationships. If an expected state transition is not connected to a task acceptance relationship and there is no task direction pointing to an adjacent device in the digital twin, retain the corresponding boundary of the expected state transition as a single-device transition boundary. The single-device transition boundary records the state changes within the same device and does not write adjacent device acceptance relationships.

[0028] Read the boundaries where task succession relationships already exist but have not passed the succession check. If the state after the previous device migration is not a completed state, the state before the next device migration is not an entering state, the state after the previous device migration occurs after the state before the next device migration, or the task directions corresponding to the two states are opposite, it is determined that the completed state of the previous device cannot be followed by the entering state of the next device. The corresponding task succession relationship is marked as a disconnected succession boundary, while preserving the correspondence between the expected state migration of the previous device, the expected state migration of the next device, and the task directions.

[0029] The desired state transitions are arranged in the order of their occurrence and connected between adjacent desired state transitions according to their boundary categories, forming a runtime transition chain. Specifically, the arrangement of desired state transitions in the order of their occurrence and their connection between adjacent desired state transitions according to their boundary categories is as follows: Read the execution sequence of each desired state transition and arrange the desired state transitions within the same device in ascending order of occurrence time. When occurrence times are the same, the order is determined by the digital twin write time. When there is a collaborative acceptance boundary between different devices, the desired state transition of the preceding device is placed before the desired state transition of the following device. When there is a disconnection acceptance boundary between different devices, retain the task direction and connect the disconnection acceptance boundary between the corresponding two desired state transitions.

[0030] The expected state migrations are categorized into several types: those between adjacent expected state migrations where an unclosed migration boundary accesses a state gap; those between consecutive expected state migrations of the same device accessing a single device migration boundary; those between expected state migrations of different devices that can be connected through a collaborative acceptance boundary access; and those between expected state migrations of different devices that cannot be connected through a disconnected acceptance boundary access but have a task acceptance relationship. After access is completed, the expected state migrations and various boundaries together form an operational migration chain.

[0031] In this embodiment, forming the migration zone carrying actual operational support includes: Extract the device affiliation and transmission sequence from the actual measured data transmitted back from the devices. Arrange the state changes of the same device within adjacent transmission times according to their occurrence order and attach them to the same device migration segment in the operational migration chain. Specifically, this involves: The system reads the device code, transmission time, and transmission status from the device's measured feedback content. When the device code matches the device identity in the digital twin, the measured feedback content is assigned to the corresponding device's feedback set; if the device code is missing, the corresponding feedback content is assigned to the unassigned feedback set and is not included in the same device migration segment. The transmission time is the time when the device's measured feedback content enters the IoT platform for collection, and the feedback status is one of the actual operating states of the physical device, including contact, current, temperature, on / off status, occupied status, or access status.

[0032] The feedback sets from the same device are arranged in ascending order of feedback time. When feedback times are the same, the order in which they are received by the IoT platform determines their sequence. When two adjacent feedback states are inconsistent, the change between the previous and subsequent feedback states is recorded as a state change; when two adjacent feedback states are consistent, no state change is formed, and only the subsequent feedback state is retained as a stable state record. The device identity and expected state migration in the running migration chain are read, and state changes with consistent device codes are attached to the same device migration segment to form a sequence of measured changes for the same device with the feedback occurrence order.

[0033] The measured feedback content of the equipment within the same equipment migration segment is cut along the migration boundary in the running migration chain. The feedback state before the migration boundary is left on the front side of the migration, and the feedback state after the migration boundary is left on the back side of the migration, while preserving the order of occurrence before and after the migration boundary, thus forming the boundary feedback content. Specifically, the measured feedback content of the equipment within the same equipment migration segment is cut along the migration boundary in the running migration chain as follows: Read the migration boundary within the same device migration segment. The migration boundary is the dividing position between the pre-migration state and the post-migration state in the desired state migration. In the measured change sequence of the same device, return states occurring before the migration boundary are classified as pre-migration states, and return states occurring after the migration boundary are classified as post-migration states. When the return state and the migration boundary coincide in sequence, read the corresponding state action. If the state action points to the pre-migration state, it is classified as pre-migration state; if the state action points to the post-migration state, it is classified as post-migration state.

[0034] Before migration, at least one callback state closest to the migration boundary must be retained; after migration, at least one callback state closest to the migration boundary must also be retained. If a callback state is missing before migration, the missing position is marked as a front-side callback gap; if a callback state is missing after migration, the missing position is marked as a rear-side callback gap. The front-side callback states, the rear-side callback states, the callback gaps, and their order of occurrence are merged into boundary callback content, which is retained at the corresponding migration boundary.

[0035] Verify the state changes in the boundary feedback content against the expected states on both sides of the migration boundary. The state changes must be able to be cut into running support pieces when transitioning from the pre-migration state to the post-migration state. Specifically, the state changes must be able to be cut into running support pieces when transitioning from the pre-migration state to the post-migration state. Read the pre-migration and post-migration return states from the boundary return data, and read the pre-migration and post-migration states on both sides of the migration boundary. Discrete states are checked using state-action type verification. If the action type of the pre-migration return state is consistent with the pre-migration state, and the action type of the post-migration return state is consistent with the post-migration state, and the pre-migration return state occurs before the post-migration return state, then the discrete state change can be determined to be continuous.

[0036] Continuous numerical states are checked using normalized difference. The returned state value and the expected state value are read. The absolute value of the difference between the returned state value and the expected state value is calculated, and then divided by the difference between the upper and lower limits of the allowable range of variation for the expected state value to obtain the normalized difference. If the normalized difference does not exceed 0.05, the continuous numerical states are considered consistent. When the allowable range of variation for the expected state value is missing, the rated operating range of similar equipment in the digital twin is used as the allowable range. When the rated operating range is missing, the corresponding continuous numerical state is classified as a source of support gap.

[0037] When state changes can be sequentially transmitted, the pre-migration and post-migration states, along with their order of occurrence, are extracted from the boundary feedback content to form a running support piece. If multiple sets of sequentially transmitted state changes exist within the boundary feedback content, the time interval between the pre-migration feedback state and the migration boundary, and the time interval between the post-migration feedback state and the migration boundary are calculated for each set of state changes. The set of state changes with the smallest sum of these two time intervals is selected as the running support piece, while the remaining state changes are retained as reference records for feedback from the same device.

[0038] When a state change cannot be seamlessly transitioned from the pre-migration state to the post-migration state, the boundary backhaul content is retained at the migration boundary and marked as a support gap. Specifically, retaining the boundary backhaul content at the migration boundary and marking it as a support gap involves: Read the pre-migration and post-migration return status from the boundary return content. If the pre-migration return status is missing, the post-migration return status is missing, the pre-migration return status is inconsistent with the pre-migration status, the post-migration return status is inconsistent with the post-migration status, or the post-migration return status is earlier than the pre-migration return status, it is determined that the state change cannot be continued from the pre-migration status to the post-migration status.

[0039] When a state change cannot be continued, the running support piece is not clipped, and the boundary feedback content is retained at the corresponding migration boundary. A missing feedback state before migration is marked as a front support gap; a missing feedback state after migration is marked as a rear support gap; inconsistent states are marked as state misalignment support gaps; and reversed order is marked as a sequence reversal support gap. Support gaps retain the corresponding device code, migration boundary position, and the source of the unsuccessful feedback state.

[0040] The running support piece is pressed onto the corresponding migration boundary, and the support notch is left at the corresponding migration boundary, thus converging to form a migration range carrying the actual running support. Specifically, the converging to form a migration range carrying the actual running support is as follows: Read the running support piece and support gap corresponding to each migration boundary. When a running support piece exists, align the front migration feedback state in the running support piece with the front of the migration boundary, align the rear migration feedback state in the running support piece with the rear of the migration boundary, and attach the order of occurrence to the migration boundary to form a migration boundary with real running support.

[0041] When a support gap exists, it is retained at the migration boundary and does not replace the running support piece; the location of the support gap is retained as the source of the support gap and maintains a corresponding relationship with the migration boundary. When both running support pieces and support gaps exist at the migration boundary, the running support piece is used as the state to continue support, and the support gap is used as the source of boundary anomalies. The running support piece and the support gap are retained at different locations on the migration boundary.

[0042] The expected state migration, operating support piece, and support gap at the same migration boundary are converged into a migration interval. The migration interval must at least include the expected state migration and the migration boundary; when an operating support piece exists, the migration interval carries the actual operating support; when a support gap exists, the migration interval carries the source of the support gap. After convergence, the migration interval carrying the actual operating support is used as the final result after aligning the equipment's measured feedback content with the operating migration chain.

[0043] In this embodiment, the improved neural controlled differential equation model includes a control path, an initialization network, a vector field network, a differential equation solver, an interval initial value update layer, and a readout layer. The interval initial value update layer is embedded at the intersection of adjacent integration intervals of the differential equation solver. The improved structure and parameter acquisition method of the neural controlled differential equation model are as follows: When processing irregular time series, the ordinary neural controlled differential equation model interpolates the discrete observation sequence into a continuous control path. The initial network maps the initial observation to the entry hidden state. The vector field network drives the hidden state to continuously integrate along the control path in the differential equation solver. The readout layer reads the hidden state and outputs the judgment result. At the junction of adjacent integration intervals, the ordinary neural controlled differential equation model adopts the method of directly accepting the exit hidden state as the initial value of the next integration interval. Boundary disconnection information such as device acknowledgment completion but entity state unchanged, entity state change but digital twin state not transferred, and previous device completion but adjacent device not accepted is absorbed by the continuous integration process, resulting in the running breakpoints being smoothed into normal state changes.

[0044] Improved neural controlled differential equation models include: Control path: It inherits the sequence of state changes within the migration interval and organizes the changes between the state before migration and the state after migration into a continuous control segment; Initial network: Read the pre-transition state of the first integration interval and map the pre-transition state to the entry hidden state; Vector field network: Reads the hidden state of the entry point and continuous control segments, and converts the control path increment in the continuous control segments into interval propulsion. Differential equation solver: Accumulates interval advance along continuous control segments to form the exit hidden state corresponding to each integration interval; Interval Initial Value Update Layer: At the junction of adjacent integration intervals embedded in the differential equation solver, the interface for directly writing the initial value of the next integration interval from the hidden exit state is cut off. Readout layer: Reads the hidden state chain carrying the initial value components and outputs the runtime evaluation results.

[0045] Before model deployment, the model is trained and calibrated using historical operational migration chains, actual equipment measurement feedback, and manually verified operational disconnection annotation samples. The linear mapping matrix and bias in the initial network, the multi-layer mapping parameters in the vector field network, and the classification mapping parameters in the readout layer are learnable parameters, updated in reverse using migration intervals and operational disconnection annotations from the training samples. The initial value carry-over mapping parameters and interval change write-to-mapping mapping parameters in the interval initial value update layer are also learnable parameters, learning the correspondence between the entry hidden state, the exit hidden state, and the control path increment. The state consistency threshold, feedback time alignment tolerance, and continuous numerical state normalization difference threshold are engineering judgment thresholds, calibrated using historical operational data from similar equipment before deployment; in one implementation, the continuous numerical state normalization difference threshold is set to 0.05, and the feedback time alignment tolerance is set to half the corresponding equipment sampling period. When training annotations are lacking, the initial network and vector field network are trained using closed migration records within a digital twin, and then the interval initial value update layer and readout layer are calibrated using manually verified abnormal boundary records.

[0046] The improved model adds an interval initial value update layer inside the solver of the ordinary neural controlled differential equation model. The original continuous integration path no longer passes the hidden state of the exit of the previous integration interval to the initial value entry of the next integration interval as is. Instead, it reorganizes the source of initial values ​​at the junction of adjacent integration intervals. After this structural modification, the model retains the unclosed changes at the migration boundary, providing an internal state basis for the formation of closed initial value components, retained initial value components, and runtime evaluation results.

[0047] In this embodiment, forming an outlet concealment state includes: The migration intervals carrying actual operational support are arranged into adjacent integration intervals according to the occurrence order of the operational migration chain, and the migration boundaries between adjacent integration intervals are aligned as integration handover boundaries. Specifically, arranging the migration intervals carrying actual operational support into adjacent integration intervals according to the occurrence order of the operational migration chain is as follows: The expected state transition, running support piece, and transition boundary are read from the transition interval. The expected state transition provides the start and end range of the state in the integration interval, the running support piece provides the physical device's feedback state, and the transition boundary provides the handover position of the integration interval. The occurrence order of the running transition chain is jointly determined by the occurrence time of the expected state transition and the task direction of the collaborative acceptance boundary. Within the same device, they are arranged in ascending order of occurrence time, and between different devices, they are arranged in order of task direction from the previous device to the next device.

[0048] Each migration interval after arrangement is converted into an integration interval. When the migration boundary of the previous migration interval has a continuity relationship with the migration boundary of the next migration interval, the two migration boundaries are aligned as integration junction boundaries. When a migration interval lacks a migration boundary, the migration interval enters the set to be aligned and does not participate in the current round of integration interval arrangement; the set to be aligned retains the device identity and expected state migration, which is used for rearrangement after the running migration chain is corrected.

[0049] The control path intercepts the state change sequence within each integral interval along the integral junction boundary, compressing the continuous change from the pre-transition state to the post-transition state into control path increments, and forming consecutive control segments in the order of adjacent integral intervals. Specifically, the composition of the control path and the compression of the continuous change from the pre-transition state to the post-transition state into control path increments are as follows: The control path consists of a status reading unit, an incremental encoding unit, a time interval encoding unit, and a default channel filling unit.

[0050] The status reading unit reads the pre-migration status, post-migration status, and running support chip within the integration interval, converting discrete states into status action codes and continuous numerical states into normalized status values. The normalized status value is calculated by subtracting the lower limit of the rated range from the read status value, and then dividing by the difference between the upper and lower limits of the rated range for similar equipment. If the rated range is missing, the minimum and maximum values ​​of historical stable operating data for similar equipment are used as the rated range; if historical stable operating data is missing, continuous numerical states enter the default channel, and the default channel filling unit writes 0.

[0051] The incremental encoding unit encodes the state actions and normalized state value changes between the pre-migration and post-migration states into state change increments. The time interval encoding unit reads the return time interval between the pre-migration and post-migration states and reads the upper limit of the sampling period for similar devices. It divides the return time interval by the upper limit of the sampling period for similar devices to obtain the time interval normalized value. When the time interval normalized value is less than 0, it writes 0; when the time interval normalized value is greater than 1, it writes 1; when the time interval normalized value is within the range of 0 to 1, the calculated value is retained. When the upper limit of the sampling period for similar devices is missing, the maximum interval between two adjacent return times in the historical stable operation record of similar devices is read as the upper limit of the sampling period. When the historical stable operation record is missing, the time interval encoding unit writes the time interval value into the default channel, and the default channel is filled with 0.

[0052] The state change increments, normalized feedback time intervals, and running support plate closure marks are arranged to form control path increments. In one embodiment, each control path increment is a 24-dimensional vector, with discrete state action codes occupying 8 dimensions, continuous numerical state changes occupying 8 dimensions, time intervals occupying 4 dimensions, and running support plate closure marks occupying 4 dimensions. The control path increments are arranged in the order of occurrence within the integration interval to form continuous control segments; the arrangement order of continuous control segments is consistent with that of adjacent integration intervals.

[0053] The initial network receives continuous control segments from the first integration interval and compresses the pre-transition state of the first integration interval into an entry hidden state. The composition of the initial network and the entry hidden state compression process are as follows: The initial network consists of a state vector sorting unit, a linear mapping unit, a bias superposition unit, and a amplitude limiting activation unit.

[0054] The state vector processing unit reads the pre-transition state and the first control path increment of the first integration interval, and concatenates the discrete state action codes, normalized state values, and time position codes in the pre-transition state into an initial state vector. In one embodiment, the initial state vector is 32-dimensional, the linear mapping matrix is ​​32×64, and the bias vector is 64-dimensional.

[0055] The linear mapping unit multiplies each eigenvalue in the initial state vector with the parameter of the corresponding row in the linear mapping matrix, and then accumulates column-by-column across the 64 hidden dimensions to obtain the initial hidden response. The bias stacking unit stacks the 64-dimensional bias vector onto the initial hidden response dimension-by-dimensional. The amplitude-limited activation unit performs hyperbolic tangent activation on the initial hidden response after the bias stacking, limiting the output value to the range of -1 to 1, forming a 64-dimensional entry hidden state. When the initial state vector has a default channel, the default channel participates in the multiplication and accumulation with 0, without changing the calculation of the remaining dimensions in the linear mapping matrix.

[0056] The vector field network reads the ingress hidden state and continuous control segments within the differential equation solver. It then applies the control path increments from the continuous control segments sequentially to the hidden dimensions corresponding to the ingress hidden state, forming interval propagation quantities in the same order as the control path increments. The differential equation solver accumulates these interval propagation quantities along the order of occurrence of the continuous control segments. The specific composition of the vector field network and the differential equation solver, as well as the process of accumulating these interval propagation quantities, are as follows: The vector field network consists of a hidden state mapping unit, an incremental action matrix unit, and a propulsion calculation unit; the differential equation solver consists of a control segment reading unit, a propulsion accumulation unit, and an anomaly backoff unit.

[0057] The hidden state mapping unit reads the entry hidden state, multiplies each dimension of the entry hidden state by the parameter of the corresponding row in the first linear mapping matrix, and accumulates column by column along the intermediate hidden dimension. The accumulated result is superimposed on the first bias vector and then fed into the hyperbolic tangent activation function to form the intermediate hidden response. The intermediate hidden dimension is determined by the model deployment configuration. In one embodiment, the entry hidden state is 64-dimensional, and the intermediate hidden dimension is twice the dimension of the entry hidden state. The first linear mapping matrix is ​​64×128, and the first bias vector is 128-dimensional. The first linear mapping matrix and the first bias vector are learnable parameters that are updated in reverse order by disconnecting labeled samples during model training.

[0058] The incremental action matrix unit reads the intermediate hidden response, multiplies each dimension of the intermediate hidden response by the parameter of the corresponding row in the second linear mapping matrix, and accumulates column by column at the vector field expansion position. The accumulated result is superimposed on the second bias vector to form the expanded vector. The number of vector field expansion positions is determined by multiplying the number of dimensions of the entry hidden state by the number of dimensions of the control path increment. In one embodiment, the entry hidden state is 64-dimensional, the control path increment is 24-dimensional, the expanded vector is 1536-dimensional, the second linear mapping matrix is ​​128×1536, and the second bias vector is 1536-dimensional. The expanded vector is rearranged in row order to form a 64×24 vector field matrix. The 64 rows of the matrix correspond to the 64 hidden dimensions of the entry hidden state, and the 24 columns of the matrix correspond to the 24 dimensions of the control path increment. The value at each position in the matrix represents the intensity of the effect of a control path increment dimension on a hidden dimension. When the expanded vector is non-numerical, the anomaly rollback unit rolls back the current integration position to the previous effective hidden state and marks the corresponding integration boundary as a vector field matrix generation anomaly.

[0059] The control segment reading unit reads the 24-dimensional control path increments sequentially according to the order in which the control path increments occur in the continuous control segments. The propulsion calculation unit multiplies each row of the vector field matrix with the control path increment dimension by dimension and accumulates the results to obtain a 64-dimensional interval propulsion. The propulsion accumulation unit superimposes the interval propulsion onto the current entry hidden state to form the hidden state of the current integration position. When the interval propulsion exceeds the range of -3 to 3, the interval propulsion is limited to -3 to 3; when a default channel exists for the control path increment, the default channel participates in the vector field matrix multiplication as 0. After all control path increments are accumulated, the end hidden state of the current integration interval is obtained.

[0060] The hidden entry state, after being accumulated through interval advancement, forms a corresponding hidden exit state, maintaining a correspondence with the integral boundary. Specifically, the hidden entry state, after being accumulated through interval advancement, forms a corresponding hidden exit state as follows: Read all interval advance quantities within the current integration interval, using the entry hidden state as the starting point for accumulation. Stack the interval advance quantities sequentially according to the order of control path increments. The exit hidden state is obtained after the last control path increment is stacked. The exit hidden state maintains the 64-dimensional hidden dimension order and records the integration boundary corresponding to the current integration interval.

[0061] When the exit hidden state is non-numerical, exceeds the limit range, or has an empty continuous control segment, the abnormal rollback unit is activated. Exit hidden states that are non-numerical or exceed the limit range revert to the previous valid hidden state. For integral intervals with empty continuous control segments, the entry hidden state is retained as a temporary exit hidden state, and the corresponding integral handover boundary is marked as a control path gap. After processing, the exit hidden state maintains a correspondence with the integral handover boundary, serving as the integral propagation result for the current integral interval.

[0062] In this embodiment, the initial value update object is formed as follows: The interval initial value update layer intercepts the direct write interface between the hidden exit state of the previous integration interval and the initial value entry of the next integration interval at the integration handover boundary. It closes the transmission end where the hidden exit state enters the initial value entry as is, and retains the integral correspondence between the hidden entry state of the previous integration interval and the continuous control segment. The composition of the interval initial value update layer and the direct write interface interception process are as follows: The interval initial value update layer consists of a direct write interface interception unit, an initial value receiving base construction unit, an interval change back-and-forth unit, and an initial value source reconnection unit.

[0063] The direct-write interface interception unit reads the hidden state of the previous integration interval exit formed at the integration boundary of the differential equation solver, and reads the receiving address of the initial value entry of the next integration interval. The direct-write interface writes the connection relationship of the previous integration interval exit hidden state into the initial value entry of the next integration interval as is. The previous integration interval exit hidden state is a 64-dimensional vector, and the initial value entry of the next integration interval is the position where the differential equation solver reads the initial hidden state when starting the next integration interval.

[0064] The direct-write interface interception unit rewrites the connection state between the exit hidden state and the initial value entry to a closed state. After the closed state is written to the integral handover boundary, the exit hidden state no longer enters the initial value entry of the next integral interval with the original 64-dimensional vector. The integral correspondence between the previous integral interval entry hidden state and the continuous control segment is synchronously retained. The integral correspondence records the sequential connection between the 64 hidden dimensions of the entry hidden state, the 24-dimensional control path increment in the continuous control segment, and the interval advancement formed by the cumulative connection of the vector field network.

[0065] The initial value basis is established based on the hidden state of the previous integration interval entry. The initial value basis is expanded sequentially along the hidden dimension of the entry hidden state. The initial value entry of the next integration interval is then reconnected to the initial value basis. The process of establishing the initial value basis and reconnecting the initial value entry is as follows: The initial value basis construction unit reads the hidden state of the previous integration interval and arranges the 1st to 64th dimensions of the hidden state sequentially as the initial value basis. Each hidden dimension in the initial value basis inherits the value of the corresponding dimension in the hidden state, and the order of the hidden dimensions is consistent with that of the hidden state. The hidden dimensions formed by the default channel in the hidden state retain the already calculated values ​​and are not additionally marked with default tags.

[0066] When the initial value entry point of the next integration interval is reconnected to the initial value basis, the differential equation solver switches the source of the initial value for the next integration interval from the hidden state of the previous integration interval exit to the initial value basis. Before the initial value basis completes the interval transformation and is written, the initial value entry point of the next integration interval remains in a pending read state; after the initial value basis completes dimension merging, the initial value entry point of the next integration interval reads the merged 64-dimensional basis vector. The dimension reconnection parameters in the initial value basis construction unit are learnable parameters, updated in reverse with disconnected labeled samples during the training phase, and remain fixed during the deployment phase.

[0067] The control path increments along the continuous control segment sequentially call the accumulated results formed by the vector field network in the differential equation solver. The hidden changes generated by the exit hidden state of the previous integration interval relative to the entrance hidden state of the previous integration interval are then decomposed back to the source of the control path increments according to the hidden dimension corresponding to the control path increments, forming an interval change sequence consistent with the control path increment order. The process of decomposing the hidden changes back to the source of the control path increments is specifically as follows: The interval change back-splitting unit reads the interval advance sequence stored by the differential equation solver in the previous integration interval. In each continuous control segment, the 24-dimensional control path increment trigger vector field network outputs a 64-dimensional interval advance. The differential equation solver accumulates these 64-dimensional interval advances to the current hidden state according to the order in which the control path increments occur. The interval advance sequence retains the source, order of occurrence, and corresponding hidden dimension of the control path increments.

[0068] The interval change back-splitting unit reads the hidden state of the previous integration interval's exit and entrance, and performs a dimension-by-dimensional subtraction on the same hidden dimension to form a 64-dimensional hidden change vector. The dimension-by-dimensional subtraction is calculated by subtracting the i-th dimension value of the entrance hidden state from the i-th dimension value of the exit hidden state to obtain the i-th hidden change, where i increases from 1 to 64.

[0069] The hidden change vector serves as the total amount verification object, and the interval advancement sequence serves as the change source object. The interval change splitting unit splits the 64-dimensional interval advancement triggered by each control path increment into an interval change and writes the interval change into an interval change sequence that is consistent with the order in which the control path increment occurs. When the absolute difference between the cumulative value of the interval advancement sequence on the same hidden dimension and the corresponding dimension of the hidden change vector does not exceed 0.001, the splitting is considered consistent; when the absolute difference exceeds 0.001, the corresponding hidden dimension enters the splitting anomaly set. The hidden changes in the splitting anomaly set are not written into the initial value basis but are retained at the integral boundary as the source of abnormal changes.

[0070] The interval change sequence is aligned with the same-order hidden dimensions of the initial value basis according to the control path increment order. The correspondence between the dimensions of the interval change and the pre-migration state in the initial value basis is checked item by item. The interval change sequence is then written into the initial value basis. The source of the initial value for the subsequent integration interval is changed from direct writing of the exit hidden state to receiving and superimposing the interval change from the entry hidden state. The process of writing the interval change sequence into the initial value basis is as follows: The initial value source modification unit reads each interval change in the interval change sequence, and the interval change retains the source of the control path increment and the hidden dimension position. When the hidden dimension position is consistent with the dimension index of the initial value basis, there is a dimension correspondence between the interval change and the initial value basis; when the hidden dimension position is missing or falls into the back-split anomaly set, the interval change is retained at the integral handover boundary and does not participate in this round of writing.

[0071] When a dimensional correspondence exists, the initial value source modification unit superimposes the interval change values ​​onto the base value of the corresponding hidden dimension of the initial value receiving base. The superposition calculation involves reading the i-th dimension base value of the initial value receiving base, adding the corresponding i-th dimension change value from the interval change sequence, and obtaining the i-th dimension update value. When the i-th dimension update value exceeds the range of -1 to 1, it is limited to -1 to 1. When a hidden dimension has multiple interval changes, they are superimposed sequentially according to the order of control path increments, with limiting performed after each superposition.

[0072] After the interval change sequence is written, the initial value basis retains the basis values ​​of the ingress hidden state and the interval changes extracted from the control path increment. The source of the initial value for the next integration interval changes from direct writing of the exit hidden state to receiving the ingress hidden state and superimposing the interval changes. The interval change sequence continues to retain the source relationship with the continuous control segment. The interval change writing mapping parameters in the initial value source change unit are learnable parameters, updated in reverse with disconnected labeled samples during the training phase, and remain fixed during the deployment phase.

[0073] The initial value basis after writing the interval change sequence is dimension-merged. The merged initial value basis replaces the hidden state at the exit of the previous integration interval and becomes the source of the initial value for the next integration interval. The dimension-merging process of the initial value basis is as follows: The initial value source modification unit reads the initial value basis after writing the interval change sequence, and checks the update value dimension by dimension from the 1st to the 64th dimension according to the hidden dimension order. Only one final update value is retained for each hidden dimension; when multiple interval changes are written to the same hidden dimension, after completing the item-by-item superposition and amplitude limiting according to the order of the control path increment, the value after the last amplitude limiting is used as the final update value.

[0074] After all hidden dimensions have been checked, the 64 final updated values ​​are arranged in the original order of the hidden dimensions to form the merged initial value base. The merged initial value base replaces the hidden state at the exit of the previous integration interval, becoming the source of the initial values ​​for the next integration interval. The hidden state at the exit of the previous integration interval is retained in the historical state cache of the integration boundary, serving only as a source record before interface interception and not entering the initial value entry point of the next integration interval. The merged initial value base serves as the core vector in the initial value update object, handling the process of separating closed initial value components from stagnant initial value components.

[0075] The direct-write interface from the hidden state at the exit of the previous integration interval to the initial value entry of the next integration interval is closed, forming an initial value update object carrying the initial value basis and the interval change sequence. The formation process of the initial value update object is as follows: The direct-write interface interception unit writes the connection state between the hidden state of the previous integral interval exit and the initial value entry of the next integral interval into a closing flag. The closing flag is bound to the integral junction boundary. When the closing flag exists, the reading object of the initial value entry of the next integral interval is limited to the merged initial value receiving base, and the hidden state of the previous integral interval exit is retained in the historical state cache of the integral junction boundary.

[0076] The initial value update object consists of the merged initial value basis, the interval change sequence, the set of reverted anomalies, and the integral handover boundary. The merged initial value basis records the source of the initial value for the next integral interval, the interval change sequence records the hidden changes caused by the control path increment, the set of reverted anomalies records the hidden changes for which the source of the control path increment could not be reverted, and the integral handover boundary records the closing position of the direct write interface. After the initial value update object is formed, it serves as the result of rewriting the initial value interface of adjacent integral intervals, inheriting the formation process of closed initial value components and stagnant initial value components.

[0077] Compared to the ordinary neural controlled differential equation model, the improvement of this implementation lies in transforming the interface between the integral intervals of the differential equation solver from a single state inheritance structure to a reconfigurable initial value interface. In the ordinary neural controlled differential equation model, the interface only retains the exit hidden state reading port, and the subsequent integral interval directly reads the overall exit hidden vector as the initial value. This implementation adds an initial value receiving channel, an interval change back-off channel, and an interface closure state at the interface. The solver simultaneously retains the entry basis, change source, writing dimension, and interface state at the interface position. The changes generated by the vector field network in the previous integral interval are no longer mixed into the overall exit hidden vector, but are retained within the solver as structured intermediate quantities that can locate the control path increment and hidden dimension; the subsequent integral interval reads the initial value source after basis reconstruction. After the interface structure modification, the state inheritance path and recursive boundary between adjacent integral intervals of the neural controlled differential equation model are changed. Unclosed changes on the running transition boundary have independent bearing positions within the model, and the division between closed initial value components and stagnant initial value components has a traceable internal state source.

[0078] In this embodiment, forming the closed initial value component and the lingering initial value component includes: The initial value inheriting basis and interval change sequence in the initial value update object are inherited. The interval change sequence is aligned item by item with the corresponding migration boundary running support piece along the control path increment occurrence order. Specifically, the alignment of the interval change sequence with the corresponding migration boundary running support piece along the control path increment occurrence order is as follows: The initial value update object is read from the initial value basis, interval change sequence, integral boundary, and back-split anomaly set. The initial value basis is a 64-dimensional basis vector. The interval change sequence consists of interval changes triggered by the control path increment. Each interval change retains the source of the control path increment, the hidden dimension position, and the order of occurrence. The interval changes in the back-split anomaly set are not included in the current round of closure judgment and are retained as the source of anomalies at the integral boundary.

[0079] The control path increment source corresponding to each interval change in the interval change sequence is read, and the control path increment source is bound to the migration boundary corresponding to the integral handover boundary. When a running support piece exists at the migration boundary, an alignment relationship is established between the interval change and the running support piece; when a running support piece is missing at the migration boundary, the interval change is temporarily aligned with the support gap at the migration boundary. The interval change sequence is processed item by item in the order of occurrence of control path increments. Interval changes that occur earlier are aligned with the running support piece first, and interval changes that occur later are arranged sequentially along the same migration boundary.

[0080] Using the ability of the running support piece to seamlessly transition from the pre-migration state to the post-migration state as the closure criterion, the interval changes for which the closure criterion holds are preserved in the same-order hidden dimension of the initial value's supporting basis. Specifically: The closure criterion in this step is jointly determined by the pre-migration feedback state, post-migration feedback state, and their order of occurrence in the running support plate. The discrete state reads the action type of the state. If the action type of the pre-migration feedback state is consistent with the pre-migration state, and the action type of the post-migration feedback state is consistent with the post-migration state, and the pre-migration feedback state precedes the post-migration feedback state, then the closure criterion is established.

[0081] The system continuously reads the returned status values ​​and the desired status values. It calculates the absolute value of the difference between the returned status value and the desired status value, then divides this absolute value by the difference between the upper and lower limits of the allowed range of variation for the desired status value to obtain the normalized difference. If the normalized difference does not exceed 0.05, and the returned status before migration is earlier than the returned status after migration, the closure criterion is established. If the allowed range of variation for the desired status value is missing, the rated operating range of similar equipment is read; if the rated operating range of similar equipment is missing, the continuous numerical status enters a support gap and does not form a closure criterion.

[0082] When the closure criterion is valid, the hidden dimension position of the interval change is read, and the interval change is retained in the same-order hidden dimension of the initial value's basis. When the hidden dimension corresponding to the initial value's basis has already been written into the interval change, the newly retained interval change is arranged after the existing interval changes according to the order of the control path increment, and the corresponding order mark is retained. Interval changes where the closure criterion is invalid enter the unclosed criterion verification process.

[0083] Using the absence of a continuous state in the running support plate as the basis for unclosed intervals, the interval changes where the unclosed interval basis is valid are stripped from the corresponding hidden dimension of the initial value base and returned to the migration boundary. Specifically, the interval changes where the unclosed interval basis is valid are stripped from the corresponding hidden dimension of the initial value base and returned to the migration boundary. In this step, the criteria for non-closure are determined by the absence of running support pieces, missing feedback status, inconsistent status, or reversed order of events. When running support pieces are missing from the migration boundary, the change in the alignment support gap is considered non-closure. The criteria for non-closure are established when the feedback status before migration is missing, the feedback status after migration is missing, the discrete state action types are inconsistent, the normalization difference of continuous numerical states is greater than 0.05, or the feedback status after migration is earlier than the feedback status before migration.

[0084] When the non-closure criterion is valid, the hidden dimension position of the interval change in the initial value basis is read, and the corresponding interval change is removed from the set to be merged of the hidden dimensions. The removal process retains the source of the control path increment, the hidden dimension position, and the order of occurrence of the interval change, without modifying the original basis values ​​in the initial value basis. When the interval change has already participated in the overlay, the basis values ​​and interval change values ​​before overlay are read, and the overlay result is rolled back to the basis values ​​before overlay; when there are multiple interval changes in the same hidden dimension, only the interval changes with non-closure criteria are removed, while the interval changes with closure criteria are retained.

[0085] When reverting to the migration boundary, the interval changes to be stripped are written into the retention set of the migration boundary. The retention set records the source of the interval change, the location of the hidden dimension, the order of the control path increment, and the type of unclosed basis. The type of unclosed basis corresponds to support gaps, inconsistent states, numerical deviations, or reversed order. After the revert is completed, the migration boundary obtains interval change objects that can be arranged for the retention initial value components.

[0086] The interval variations preserved within the initial value basis are merged sequentially along the hidden dimension. The merged initial value basis is then closed into closed initial value components. Specifically, closing the merged initial value basis into closed initial value components involves: The initial value is read from the interval changes already preserved in the basis through closure criteria. Hidden dimensions are processed sequentially from dimension 1 to dimension 64. For each hidden dimension, the basis value and the interval changes preserved within the corresponding hidden dimension are read. Hidden dimensions without interval changes retain the basis value; hidden dimensions with one interval change add the basis value and the interval change value; hidden dimensions with multiple interval changes accumulate them sequentially according to the order of control path increments.

[0087] After each accumulation, a limiting process is performed: values ​​exceeding 1 are written as 1, values ​​below -1 are written as -1, and values ​​between -1 and 1 are retained. After merging the 64 hidden dimensions, a 64-dimensional merged basis vector is obtained. The merged basis vector retains the basis order of the entry hidden state and the writing results of the closed interval changes, with closure defined as the closed initial value component. The closed initial value component records the corresponding integral handover boundary and the incremental order of the control path participating in the merging, serving as the initial value component that can continue to support integral advancement.

[0088] The interval changes at the retreat migration boundary are arranged sequentially along the control path increments, maintaining their correspondence with the migration boundary, forming a stagnant initial value component. Specifically, the formation of the stagnant initial value component is as follows: Read the interval changes in the migration boundary retention set and arrange them according to the order of control path increment occurrence. Interval changes with the same occurrence order are arranged in ascending order of hidden dimension number. Each interval change retains the hidden dimension position, interval change value, control path increment source, and unclosed basis type. When the interval change value exceeds the range of -1 to 1, the interval change value is limited to -1 to 1.

[0089] The arranged interval changes are encapsulated as retained initial value components. Each retained initial value component maintains a one-to-one correspondence with the migration boundary; when multiple back interval changes exist for the same migration boundary, these multiple back interval changes are merged into a single retained initial value component, internally preserving the order of control path increment occurrence. If the migration boundary lacks back interval changes, no retained initial value component is formed. After processing, the closed initial value component and the retained initial value component are respectively used as initial value update objects and verified through support closure.

[0090] Compared to the ordinary neural controlled differential equation model, this implementation inherits the internal rewriting results of the initial value interface from the interval initial value update layer, establishes closed change retention paths within the initial value inheritance base, and establishes unclosed change retention paths at the migration boundary. The ordinary neural controlled differential equation model does not distinguish between true closed and unclosed changes before the exit hidden state enters the next integration interval; this implementation separately carries these two types of changes during the initial value formation stage. The closed initial value component is read as a recursive initial value source in the subsequent integration interval, while the retained initial value component remains at the migration boundary as an independent internal state source. This processing constitutes internal state diversion within the interval initial value update layer. Combined with the reconfigurable initial value interface, it forms an internal recursive chain from interface rewriting to component attribution, providing an internal model basis for traceable readout of runtime disconnections and twin migration retention.

[0091] In this embodiment, the results of the IoT device operation evaluation include: The closed initial value component is connected to the hidden state at the entrance of the next integration interval, and the process continues along the continuous control segment of the next integration interval, forming a chain of continuous hidden states. Specifically, the closed initial value component is connected to the hidden state at the entrance of the next integration interval and the process continues: Read the 64-dimensional merged basis vector from the closed initial value component and write it into the hidden state of the next integration interval. The hidden state of the next integration interval maintains the order of the 64-dimensional hidden dimensions, which is consistent with the merged basis vector in the closed initial value component. If the closed initial value component lacks a corresponding integration boundary, the closed initial value component enters the boundary check set and does not participate in the continued advancement of this round.

[0092] The system reads the continuous control segments of the next integration interval. Each continuous control segment consists of 24-dimensional control path increments arranged in chronological order. The differential equation solver uses the entry hidden state after writing the closed initial value components as the starting point for accumulation. It reads the control path increments item by item along the continuous control segments and calls the vector field network to form a 64-dimensional interval advance. Each interval advance is superimposed on the current hidden state to form the hidden state of the current integration position. When the interval advance exceeds the range of -3 to 3, the interval advance is limited to -3 to 3. After all continuous control segments are read, the exit hidden state of the next integration interval is obtained.

[0093] The hidden states at the entrance of the next integral interval, the hidden states at each integral position within the continuous control segment, and the hidden states at the exit of the next integral interval are arranged in the order of the control path increment to form a chain of continuous hidden states. When the continuous control segment is empty, the initial value component of the closure is retained as a temporary continuous hidden state, the integral handover boundary is marked as a control segment gap, and the control segment gap enters the readout layer verification range.

[0094] The stagnant initial value component is attached to the corresponding transition boundary in the continuing hidden state chain, while maintaining the order of occurrence between the stagnant initial value component and the corresponding interval change. Specifically, the stagnant initial value component is attached to the corresponding transition boundary in the continuing hidden state chain as follows: Read the migration boundary, hidden dimension position, interval change value, control path increment source, and unclosed basis type from the retained initial value component. When the same migration boundary exists in the continuing hidden state chain, attach the retained initial value component to the same migration boundary in the continuing hidden state chain; when the continuing hidden state chain lacks the same migration boundary, keep the retained initial value component in the boundary to be attached set and record the missing migration boundary.

[0095] When multiple initial value components exist at the same migration boundary, they are arranged according to the order in which control path increments occur; if the order of control path increment occurrences is the same, they are arranged in ascending order of hidden dimension number. After arrangement, the occurrence order between the initial value components and the corresponding interval changes is written into the boundary attachment record of the continuing hidden state chain. When the interval change value exceeds the range of -1 to 1, the interval change value is limited to -1 to 1 before being entered into the boundary attachment record.

[0096] The readout layer reads the transition boundary with the lingering initial value component along the continuous hidden state chain and verifies the state continuity relationship between the running support piece at the transition boundary and the desired state transition. The specific process of verifying the composition and state continuity relationship of the readout layer is as follows: The readout layer consists of a boundary readout unit, a state continuation verification unit, a type mapping unit, and a result write-back unit.

[0097] The boundary reading unit reads the migration boundaries in the order of the continuous hidden state chain, checking whether each migration boundary is attached to a stagnant initial value component. When a migration boundary is attached to a stagnant initial value component, the boundary reading unit extracts the hidden dimension position, interval change value, and unclosed basis type from the stagnant initial value component, and simultaneously reads the running support piece and expected state migration at the migration boundary. When a migration boundary is not attached to a stagnant initial value component, the migration boundary does not enter the type mapping unit.

[0098] The state continuity verification unit reads the pre-migration and post-migration feedback states from the running support chip, and also reads the pre-migration and post-migration states in the expected state transition. Discrete states are checked for action type consistency. The action type of the pre-migration feedback state is consistent with the pre-migration state, and the action type of the post-migration feedback state is consistent with the post-migration state. Furthermore, if the pre-migration feedback state is earlier than the post-migration feedback state, the state continuity relationship is established.

[0099] Continuous numerical states are verified using normalized difference. The returned state value and the expected state value are read. The absolute value of the difference between the returned state value and the expected state value is calculated, and then divided by the difference between the upper and lower limits of the allowed range of variation for the expected state value to obtain the normalized difference. The state continuity relationship is valid if the normalized difference does not exceed 0.05, and the returned state before migration is earlier than the returned state after migration. If the allowed range of variation for the expected state value is missing, the rated operating range of similar equipment is read. If the rated operating range of similar equipment is also missing, the state continuity relationship enters a pending verification state, which is retained at the migration boundary and does not enter the type mapping unit.

[0100] When the state continuity relationship is established, the initial value component of the lingering state is labeled as the twin migration lingering type; when the state continuity relationship is not established, the initial value component of the lingering state is labeled as the actual operation disconnection type. The specific process for labeling the initial value component of the lingering state is as follows: The type mapping unit reads the verification result output by the state continuity verification unit. When the state continuity relationship is established and the migration boundary is attached to the stagnant initial value component, the stagnant initial value component is marked as the twin migration stagnant type, and the twin state migration record of the corresponding migration boundary in the digital twin is read; when the twin state migration record has not been updated from the pre-migration state to the post-migration state, the twin state not updated is written to the cause flag; when the twin state migration record has been updated from the pre-migration state to the post-migration state, the boundary stagnant pending verification is written to the cause flag, and the twin migration stagnant type is still retained at the corresponding migration boundary.

[0101] When the state continuity relationship is not established, the type mapping unit reads the source of the anomaly output by the state continuity verification unit. When a running support piece is missing, a cause for the missing support piece is written; when the pre-migration or post-migration feedback state is missing, a cause for the missing feedback state is written; when the discrete state action types are inconsistent, a cause for inconsistent action types is written; when the normalization difference of continuous numerical states is greater than 0.05, a cause for numerical deviation is written; when the post-migration feedback state is earlier than the pre-migration feedback state, a cause for reversed occurrence order is written. When at least one cause marker exists, the stagnant initial value component is marked as the actual running disconnection type; when multiple cause markers exist simultaneously, all cause markers are retained along with the actual running disconnection type.

[0102] The twin migration delay type and the actual operation disconnect type are written back to the operation migration chain in the digital twin to form the IoT device operation evaluation result. Specifically, the operation migration chain written back to the digital twin and the IoT device operation evaluation result are formed as follows: The result write-back unit reads the twin migration retention type and the actual running disconnect type, and reads the migration boundary corresponding to the type. When the migration boundary exists in the running migration chain within the digital twin, the type labeling result is written to the corresponding migration boundary; when the migration boundary lacks a corresponding position in the running migration chain, the type labeling result is added to the boundary write-back exception set, and the source of the migration boundary and the source of the retention initial value component are retained.

[0103] When writing back the type of twin migration retention, the result write-back unit marks the migration boundary as the twin state to be updated boundary and retains the verification result of the entity state continuation. When writing back the type of actual operation disconnection, the result write-back unit marks the migration boundary as the actual operation disconnection boundary and retains the source of the failure as the cause of the disconnection. After the running migration chain completes the type writing, the IoT device operation evaluation result is generated; the IoT device operation evaluation result includes the migration boundary, type calibration result, source of retention initial value component, and cause marker. The boundary write-back anomaly set is retained in the pending verification record of the digital twin and does not overwrite the already formed operation evaluation result.

[0104] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-device collaborative operation evaluation scenario in a smart warehousing park. The park includes automatic roller shutters, AGV charging piles, conveyor drive systems, sorting robotic arms, cold chain temperature-controlled cabinets, ventilation equipment, lighting controllers, access control controllers, and smoke detector linkage devices. During a complete operation cycle, the IoT platform connected 168 devices, the digital twin recorded 12,864 device status nodes, 2,860 device control tasks, and 536 adjacent device connection relationships. The traditional monitoring platform showed an online device rate of 99.1% and a task completion rate of 98.3%. However, on-site verification revealed issues such as some devices completing their tasks but not actually performing any actions, actions performed but not being synchronized by the digital twin, and subsequent devices failing to continue after the previous device completed its task.

[0105] The system first performs operational migration organization on the digital twins of IoT devices, pairing device state nodes within adjacent operational phases of the same device into state transition pairs based on their sequence of occurrence. In this experiment, the system organized 2860 state transition pairs from 12864 device state nodes, of which 2817 pairs consisted of paired states, and 43 state gaps were identified. The system closed the paired state transition pairs to define the desired state transition, while retaining the locations of state gaps as unclosed transition boundaries.

[0106] During the process of connecting adjacent devices, the system verifies the task connection relationships in the digital twin. Task connections where the completion state of the previous device can be seamlessly connected to the entry state of the next device are integrated into the corresponding expected state transitions, forming 497 collaborative connection boundaries. Task connections where the completion state of the previous device cannot be seamlessly connected to the entry state of the next device are marked as disconnected connection boundaries at 39 locations. Expected state transitions without integrated task connections are retained as single-device transition boundaries. After arrangement and boundary integration, the system forms 2860 segments of operational transition chains.

[0107] During the actual measurement feedback processing phase, the system read 38,240 pieces of actual measurement feedback content from the equipment. This content included equipment ownership, feedback sequence, feedback status, physical contact feedback, current changes, temperature control status changes, and digital twin status update records. The system arranged the status changes of the same equipment within adjacent feedback moments according to their occurrence sequence and attached them to the same equipment migration segment in the operational migration chain. Then, it cut the feedback content along the migration boundary. After processing, 2,860 sets of boundary feedback content were formed. Among these, boundary feedback content where the status change could be continued from the pre-migration state to the post-migration state was cut into 2,768 operational support pieces. Boundary feedback content where the status change could not be continued was retained at the migration boundary and marked as 92 support gaps, ultimately forming 2,860 migration intervals carrying actual operational support.

[0108] After the migration interval is integrated into the improved neural controlled differential equation model, the control path extracts the state change sequence within each integral interval along the integral junction boundary. Each integral interval is compressed into a 24-dimensional continuous control segment, which records the feedback interval, feedback delay, entity state change amplitude, twin state update offset, adjacent device acceptance interval, and boundary support state. The pre-migration state of the first integral interval is compressed into a 64-dimensional entry hidden state by the initial network. The vector field network reads the entry hidden state and the continuous control segment in the differential equation solver, and applies the control path increment successively to the hidden dimension corresponding to the entry hidden state, forming the interval propulsion.

[0109] In a task involving the transfer of a charging station from an AGV to an AGV vehicle, the digital twin records that after the charging station unlocks, the vehicle should enter the charging connection state. After the ordinary neural controlled differential equation model progresses along the continuous control segment, the average response of the hidden exit state increases from 0.34 to 0.71, and the traditional readout result classifies the task as a normal transfer. However, actual measurement feedback shows that the vehicle does not generate a charging current change at the corresponding migration boundary, and the running support plate cannot transition from the pre-migration state to the post-migration state.

[0110] This invention embeds an interval initial value update layer at the junction of adjacent integration intervals in the differential equation solver. It intercepts the direct-write interface between the hidden exit state of the previous integration interval and the initial value entry state of the next integration interval, and establishes an initial value basis based on the hidden entry state of the previous integration interval. The system sequentially calls the accumulated results formed by the vector field network within the differential equation solver along the control path increment of a continuous control segment. The hidden changes generated by the exit hidden state relative to the entry hidden state are then decomposed back to the corresponding control path increment source, forming an interval change sequence. In this task, the interval changes are mainly concentrated in the 16th to 29th hidden dimensions, with an average amplitude of 0.46.

[0111] After the support plate is checked, the interval changes corresponding to the unlocking of the charging pile are retained within the initial value receiving base, while the interval changes corresponding to the unconnected vehicle are stripped from the initial value receiving base and retreat to the migration boundary, forming a stagnant initial value component. Within a complete operating cycle, the system forms a total of 2729 closed initial value components and 131 stagnant initial value components, of which 92 originate from gaps in the physical operation support and 39 from disconnections at the collaborative receiving boundary. After the closed initial value component is connected, it continues to advance within the next integral interval, forming a continuous hidden state chain, and the stagnant initial value component is attached to the corresponding migration boundary.

[0112] The readout layer reads the migration boundary with the initial value component of the lingering state along the continuous hidden state chain and verifies the state continuity relationship between the running support piece and the expected state migration. Ultimately, the system identified 92 actual operational disconnections, 39 collaborative acceptance anomalies, and 24 twin migration lingering incidents. After verification using on-site maintenance logs, equipment current curves, and door magnetic contact records, the actual operational disconnections were correctly identified 87 times (94.6% accuracy); collaborative acceptance anomalies were correctly identified 36 times (92.3% accuracy); and twin migration lingering incidents were correctly identified 23 times (95.8% accuracy).

[0113] In the comparative experiment, the traditional receipt judgment method identified only 58 anomalies, with a false negative rate of 62.6%; the ordinary neural controlled differential equation model identified 101 anomalies, with a false negative rate of 34.8%; the present invention identified 155 anomalies, of which 146 were confirmed after review, with an overall accuracy rate of 94.2%, a false negative rate reduced to 5.8%, and an average anomaly localization time shortened from 16.7 minutes for the traditional method to 2.9 minutes. This embodiment demonstrates that the present invention can preserve unclosed changes at migration boundaries even after the device receipt display is complete, preventing continuous hidden state advancement from disrupting actual operation, causing twin migration stagnation, and smoothing out collaborative acceptance anomalies into normal state changes. This verifies the engineering feasibility of the method in multi-device collaborative operation evaluation scenarios.

[0114] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for evaluating the operation of IoT devices based on digital twins, characterized in that, Includes the following steps: The operation migration of digital twins of IoT devices is organized to align the expected state migration with the collaborative acceptance boundary, forming an operation migration chain; The actual measured data transmitted back from the equipment is divided into operation support pieces along the operation migration chain, and the migration boundaries are pressed together to form a migration range carrying real operation support. The migration interval access improves the neural controlled differential equation model by embedding an interval initial value update layer at the intersection of adjacent integration intervals of the differential equation solver. The migration intervals are arranged as adjacent integration intervals along the running migration chain. Each integration interval takes over the hidden entry state and advances along the interval state change to form the hidden exit state. At the junction of adjacent integration intervals, the original interface of the next integration interval initial value is truncated after the exit hidden state is directly written. The initial value is reset with the entry hidden state as the base, and the change of the exit hidden state relative to the entry hidden state is pushed into the initial value base to form the initial value update object. For changes in the initial value update object that are closed with the running support piece, the base is retained; for changes that are not closed with the running support piece, the boundary is rolled back, forming closed initial value components and retained initial value components. After the initial value component is connected to the integration interval, the advancing state and the stagnant initial value component are merged into a hidden state chain. The boundary assignment of the stagnant initial value component is marked and written back to the digital twin to form the operation evaluation result of the Internet of Things device.

2. The method for evaluating the operation of IoT devices based on digital twins according to claim 1, characterized in that, The formation of the operational migration chain includes: Extract device state nodes with state transition tags from the digital twin of IoT devices, and pair device state nodes of the same device in adjacent operation phases according to the order of occurrence. The verification criteria are whether the states before and after the migration appear in pairs. The state migration pairs that appear in pairs are closed as the expected state migration, and the positions where the state gaps are located are retained as unclosed migration boundaries. In the digital twin, the task succession relationship between adjacent devices is checked, and the task succession relationship that the completed state of the previous device can continue the state of the next device is connected between the corresponding expected state transitions to form a collaborative succession boundary. The expected state transition of the unconnected task acceptance relationship is retained as the single device migration boundary, and the task acceptance relationship that cannot be continued by the state of the previous device is marked as the disconnection acceptance boundary. The desired state transitions are arranged in the order in which they occur and connected between adjacent desired state transitions according to their boundary categories, forming a running transition chain.

3. The method for evaluating the operation of IoT devices based on digital twins according to claim 1, characterized in that, The formation of the migration interval carrying real operational support includes: Extract the device affiliation and transmission sequence from the actual measured data transmitted back from the device. Arrange the state changes of the same device in adjacent transmission times according to the order of occurrence and attach them to the same device migration segment in the running migration chain. Cut the measured feedback content of the equipment in the same equipment migration segment along the migration boundary in the running migration chain, leave the feedback state before the migration boundary on the front side of the migration, leave the feedback state after the migration boundary on the back side of the migration, and retain the order of occurrence before and after the migration boundary to form the boundary feedback content. Verify the state changes in the boundary feedback content with the expected states on both sides of the migration boundary. When the state changes can be cut into running support pieces from the state before migration to the state after migration. When a state change cannot be continued from the pre-migration state to the post-migration state, the boundary backhaul content is retained at the migration boundary and marked as a support gap. The running support piece is pressed onto the corresponding migration boundary, and the support gap is left at the corresponding migration boundary, thus forming a migration range carrying the actual running support.

4. The method for evaluating the operation of IoT devices based on digital twins according to claim 1, characterized in that, The improved neural controlled differential equation model includes a control path, an initial network, a vector field network, a differential equation solver, an interval initial value update layer, and a readout layer. The interval initial value update layer is embedded at the intersection of adjacent integration intervals of the differential equation solver.

5. The method for evaluating the operation of IoT devices based on digital twins according to claim 1, characterized in that, The formation of the hidden exit state includes: The migration intervals carrying real operational support are arranged into adjacent integration intervals along the occurrence order of the operational migration chain, and the migration boundaries between adjacent integration intervals are aligned as integration handover boundaries. The control path intercepts the state change sequence within each integral interval along the integral junction boundary, compresses the continuous change from the state before the migration to the state after the migration into the control path increment, and forms a continuous control segment with the same sequence according to the arrangement order of adjacent integral intervals. The initial network receives continuous control segments of the first integration interval and compresses the pre-transition state of the first integration interval into an entry hidden state. The vector field network reads the entry hidden state and continuous control segment within the differential equation solver, and applies the control path increment in the continuous control segment to the hidden dimension corresponding to the entry hidden state in turn, forming an interval advance quantity in the same order as the control path increment. The differential equation solver accumulates the interval advance quantity along the occurrence order of the continuous control segment. The hidden state at the entrance is accumulated through the interval advancement to form the corresponding hidden state at the exit, and maintains a corresponding relationship with the integral intersection boundary.

6. The method for evaluating the operation of IoT devices based on digital twins according to claim 1, characterized in that, The initial value update object includes: The interval initial value update layer intercepts the direct write interface between the previous integral interval exit hidden state and the next integral interval initial value entry at the integral handover boundary, closes the transmission end of the exit hidden state into the initial value entry as is, and retains the integral correspondence between the previous integral interval entry hidden state and the continuous control segment. The initial value basis is established based on the hidden state of the previous integration interval entry. The initial value basis is expanded sequentially along the hidden dimension of the entry hidden state. The initial value entry of the next integration interval is then connected to the initial value basis. The control path increments along the continuous control segment sequentially call the cumulative results formed by the vector field network in the differential equation solver. The hidden changes generated by the hidden state at the exit of the previous integral interval relative to the hidden state at the entrance of the previous integral interval are decomposed back to the source of the control path increment according to the hidden dimension corresponding to the control path increment, forming an interval change sequence consistent with the control path increment sequence. Align the interval change sequence with the same hidden dimension of the initial value receiving base in the order of control path increment, check the dimension correspondence between the interval change and the state before migration in the initial value receiving base item by item, and write the interval change sequence into the initial value receiving base. The source of the initial value of the next integration interval is changed from direct writing of the exit hidden state to receiving and superimposing the interval change in the entry hidden state. The initial value basis after writing the interval change sequence is dimension-merged. The merged initial value basis replaces the hidden state of the previous integration interval exit and becomes the source of the initial value for the next integration interval. The direct write interface from the hidden state of the previous integration interval exit to the initial value entry of the next integration interval is closed, forming an initial value update object carrying the initial value basis and the interval change sequence.

7. The method for evaluating the operation of IoT devices based on digital twins according to claim 1, characterized in that, The formation of closed initial value components and lingering initial value components includes: The initial value inheriting base and interval change sequence in the initial value update object are inherited, and the interval change sequence is aligned with the corresponding migration boundary running support piece item by item along the control path incremental occurrence order; Using the ability of the running support piece to continue from the state before migration to the state after migration as the closure criterion, the interval changes where the closure criterion is valid are retained in the same-order hidden dimension of the initial value basis. Using the non-formed state of the running support piece as the basis for non-closure, the interval change where the non-closure basis is established is stripped from the corresponding hidden dimension of the initial value receiving base and returned to the migration boundary. The interval changes retained within the initial value basis are merged sequentially along the hidden dimension, and the merged initial value basis is closed into closed initial value components. The interval changes of the retreat migration boundary are arranged sequentially along the control path increments and maintain their correspondence with the migration boundary, forming the stagnant initial value component.

8. The method for evaluating the operation of IoT devices based on digital twins according to claim 1, characterized in that, The evaluation results of the IoT device operation include: The closed initial value component is connected to the hidden state at the entrance of the next integral interval, and the process continues along the continuous control segment of the next integral interval to form a chain of continuous hidden states. The stagnant initial value component is attached to the corresponding transition boundary in the continuing hidden state chain, while maintaining the order of occurrence between the stagnant initial value component and the corresponding interval change; The readout layer reads the transition boundary with the lingering initial value component along the continuous hidden state chain and verifies the state continuity relationship between the running support piece at the transition boundary and the desired state transition. When the state continuity relationship is established, the initial value component of the lingering state is marked as the twin migration lingering state; when the state continuity relationship is not established, the initial value component of the lingering state is marked as the actual operation disconnection state. The twin migration retention type and the real operation disconnect type are written back to the operation migration chain in the digital twin to form the operation evaluation results of IoT devices.

Images