Water gate operation state monitoring method and device, storage medium and electronic equipment

Abstract

This disclosure relates to the field of water conservancy engineering technology, specifically to a method, device, storage medium, and electronic equipment for monitoring the operating status of a sluice gate. The sluice gate operating status monitoring method includes: upon detecting the opening and closing action of the sluice gate, extracting a single operating segment from continuous monitoring data and constructing an original three-phase current sequence based on the single operating segment; decoupling the original three-phase current sequence to obtain an impact-sensitive feature sequence, and constructing a feature matrix based on the impact-sensitive feature sequence; inputting the feature matrix into a trained state recognition model to perform multi-scale pseudo-temporal coding and category prototype constraint classification using the state recognition model, obtaining the predicted probability of each state category output by the state recognition model; and taking the state category corresponding to the maximum predicted probability as the current operating status of the sluice gate. The sluice gate operating status monitoring method provided by this disclosure can improve the accuracy of sluice gate operating status identification.

Description

Technical Field

[0001] This disclosure relates to the field of water conservancy engineering technology, specifically to a method for monitoring the operation status of a sluice gate, a device for monitoring the operation status of a sluice gate, a storage medium, and electronic equipment. Background Technology

[0002] As a core control infrastructure in flood control, drainage, water resource allocation, and navigation scheduling, the reliability and safety of sluice gates directly affect the flood control safety of the basin and the realization of engineering benefits. With the advancement of water conservancy informatization, sluice gate automated monitoring systems have generally acquired the ability to continuously collect electrical parameters such as the current and voltage of the gate opening and closing motors.

[0003] In actual engineering operation and maintenance, sluice gates are used in harsh environments with high humidity and variable loads for a long time. The gate structure is prone to hidden mechanical defects such as uneven wear of guide rails, jamming on one side, and abnormally increased frictional resistance of the water-stop rubber. Sudden collisions or foreign objects blocking the gate during the opening and closing process may cause serious accidents such as damage to the gate structure or even breakage of the wire rope.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this disclosure is to provide a method, device, storage medium, and electronic equipment for monitoring the operating status of a sluice gate, aiming to improve the accuracy of identifying the operating status of the sluice gate.

[0006] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part from practice of this disclosure.

[0007] According to one aspect of this disclosure, a method for monitoring the operational status of a sluice gate is provided, comprising: When the opening and closing action of the sluice gate is detected, a single operation segment is extracted from the continuous monitoring data, and the original three-phase current sequence is constructed based on the single operation segment; The original three-phase current sequence is decoupled to obtain the impact-sensitive characteristic sequence, and a feature matrix is ​​constructed based on the impact-sensitive characteristic sequence; the impact-sensitive characteristic sequence includes one or more of the following: instantaneous unbalance degree sequence, equivalent positive sequence current amplitude sequence, positive sequence current low-frequency trend sequence, and total impact intensity sequence; The feature matrix is ​​input into the trained state recognition model to perform multi-scale pseudo-temporal coding and category prototype constraint classification, thereby obtaining the predicted probability of each state category output by the state recognition model. The state category corresponding to the maximum predicted probability is taken as the current sluice gate operating state.

[0008] Optionally, the step of extracting a single running segment from continuous monitoring data includes: Based on the continuous monitoring data, the three-phase currents at each sampling time are extracted, and the value with the largest absolute value among the three-phase currents is used as the judgment value of the sampling time; the three-phase currents include the first phase current, the second phase current, and the third phase current; The sampling time when the judgment value first exceeds the start threshold is marked as the start time; Starting from the start time, the sampling time when the judgment value is continuously lower than the stop threshold for a preset duration is marked as the end time; Several sampling times are reserved before the start time and several sampling times are reserved after the end time to capture the single running segment.

[0009] Optionally, the method further includes: Based on the average value of the three-phase current at all sampling times during the single operation segment, zero offset is performed on the single operation segment; and / or The single running segment is linearly resampled along the time axis, and the linearly resampled single running segment includes a preset number of sampling times.

[0010] Optionally, when the impact-sensitive characteristic sequence includes an instantaneous imbalance sequence, the step of decoupling the original three-phase current sequence to obtain the impact-sensitive characteristic sequence includes: Based on the instantaneous current value sequence of each phase in the original three-phase current sequence, the sliding effective value is calculated according to the sliding window to obtain the amplitude sequence of each phase; For each sampling time, the average value of the three-phase amplitude is calculated based on the amplitude sequence of each phase as the average value of the three-phase common load water. The standard deviation of the three-phase amplitude relative to the average value of the three-phase common load level is calculated as the three-phase common load level dispersion. Divide the three-phase common load level dispersion by the three-phase common load level average to obtain the instantaneous unbalance value corresponding to the sampling time; The sampling times of the original three-phase current sequence are traversed to obtain the instantaneous unbalance sequence composed of the instantaneous unbalance values.

[0011] Optionally, when the impact-sensitive characteristic sequence includes a total impact intensity sequence, the step of decoupling the original three-phase current sequence to obtain the impact-sensitive characteristic sequence includes: Based on the instantaneous current value sequence of each phase in the original three-phase current sequence, the instantaneous change intensity value of each sampling time compared with the adjacent sampling time is calculated to obtain the instantaneous change intensity sequence of each phase. The local baseline is calculated based on the instantaneous change intensity sequence of each phase, and the effective impact change value of each phase is obtained by subtracting the local baseline from the instantaneous change intensity value of each phase. The effective impact change value of each phase is fused into three phases to obtain the total impact intensity sequence.

[0012] Optionally, the method further includes: pre-training the state recognition model; the pre-training of the state recognition model includes: For a single running sample in the running sample set, the impact-sensitive feature sequence of the running sample is extracted to construct a feature matrix, and an initial input feature map is constructed based on the feature matrix; The input feature map is encoded using a convolutional network to obtain an encoded output feature map; the convolutional network includes multiple convolutional blocks of different scales, and a gating network is used to adaptively generate the scale weights of each convolutional block for feature fusion. The predicted probability of each state category is calculated based on the encoded output feature map, and the total loss value is calculated based on the predicted probability. The model parameters of the convolutional network are adjusted using the total loss value until training is complete and the state recognition model is obtained.

[0013] Optionally, the total loss value includes a class-weighted cross-entropy loss term and a prototype contrast constraint loss term. The step of calculating the predicted probability of each state class based on the encoded output feature map and calculating the total loss value based on the predicted probability includes: Global average pooling is performed on the encoded output feature map along the pseudo-temporal direction to obtain the sample embedding vector; Calculate the similarity between the sample embedding vector and the category prototype vector corresponding to each state category to obtain the predicted probability of each state category; The category-weighted cross-entropy loss term is calculated based on the predicted probability of each state category; and the prototype comparison constraint loss term is calculated based on the predicted state category corresponding to the maximum predicted probability and the actual state category of the running sample.

[0014] According to a second aspect of this disclosure, a sluice gate operation status monitoring device is provided, comprising: The detection module is used to extract a single operation segment from continuous monitoring data when the opening and closing action of the sluice gate is detected, and to construct the original three-phase current sequence based on the single operation segment; A decoupling module is used to decouple the original three-phase current sequence to obtain an impact-sensitive characteristic sequence, and to construct a feature matrix based on the impact-sensitive characteristic sequence; the impact-sensitive characteristic sequence includes one or more of the following: instantaneous unbalance degree sequence, equivalent positive sequence current amplitude sequence, positive sequence current low-frequency trend sequence, and total impact intensity sequence; The prediction module is used to input the feature matrix into the trained state recognition model, so as to use the state recognition model to perform multi-scale pseudo-temporal coding and category prototype constraint classification, and obtain the predicted probability of each state category output by the state recognition model. The output module is used to take the state category corresponding to the maximum predicted probability as the current sluice gate operating state.

[0015] According to a third aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the sluice gate operation status monitoring method as described in the above embodiments.

[0016] According to a fourth aspect of the present disclosure, an electronic device is provided, characterized in that it includes: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the sluice gate operation status monitoring method as described in the above embodiments.

[0017] The exemplary embodiments disclosed herein may have some or all of the following beneficial effects: In the technical solutions provided by some embodiments of this disclosure, on the one hand, an adaptive start-stop interception mechanism is used to extract single-run segments from continuous monitoring data to construct the original three-phase current sequence, providing a standardized three-phase synchronous input for subsequent analysis and avoiding the problem of inconsistent boundaries caused by different start-stop durations and inconsistencies in start and stop times. On the other hand, a physical mechanism-driven feature decoupling system is constructed to decouple the original three-phase current sequence to obtain one or more of the following impact-sensitive feature sequences: instantaneous imbalance sequence, equivalent positive sequence current amplitude sequence, positive sequence current low-frequency trend sequence, and total impact intensity sequence. The mixed current signal with power frequency as the carrier is decomposed into independent physical channels such as structural imbalance, load trend, and transient impact, extracting a more accurate feature matrix and improving the accuracy of state monitoring. Furthermore, a lightweight state recognition model jointly optimized by multi-scale pseudo-temporal convolution and category prototype comparison constraints is introduced, taking into account both local impacts and cross-segment continuous anomalies, and forcing similar samples to cluster compactly in the embedding space. Under small sample conditions, the clarity of the distinction between slight jamming and normal heavy load is improved, further enhancing the accuracy of state category recognition.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings: Figure 1 The schematic diagram illustrates a flowchart of a sluice gate operation status monitoring method according to an exemplary embodiment of the present disclosure; Figure 2 A schematic diagram illustrating a primitive three-phase current sequence in an exemplary embodiment of the present disclosure is shown. Figure 3 This diagram schematically illustrates a three-phase amplitude sequence according to an exemplary embodiment of the present disclosure; Figure 4 This illustration schematically depicts a sequence of instantaneous imbalances in an exemplary embodiment of the present disclosure. Figure 5 This illustration schematically depicts an equivalent positive sequence current amplitude and low-frequency trend sequence in an exemplary embodiment of the present disclosure. Figure 6 This schematic diagram illustrates a total impact intensity sequence in an exemplary embodiment of the present disclosure; Figure 7 This schematic diagram illustrates the composition of a sluice gate operation status monitoring device according to an exemplary embodiment of the present disclosure; Figure 8 The schematic diagram illustrates the structure of a computer system of an electronic device according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be more thorough and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art.

[0021] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0022] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0023] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0024] As a core control infrastructure in flood control, drainage, water resource allocation, and navigation scheduling, the reliability and safety of sluice gates directly affect the flood control safety of the basin and the realization of engineering benefits. With the advancement of water conservancy informatization, sluice gate automated monitoring systems have generally acquired the ability to continuously collect electrical parameters such as the current and voltage of the gate opening and closing motors.

[0025] However, most monitoring systems are still in the rudimentary stage of threshold over-limit alarms, that is, they only trigger shutdown or alarm when the current amplitude exceeds the preset overload protection setting. This crude monitoring method is difficult to detect signs of mechanical performance degradation in the early stages of equipment operation.

[0026] In actual engineering operation and maintenance, sluice gates are used in harsh environments with high humidity and variable loads for a long time. The gate structure is prone to hidden mechanical defects such as uneven wear of guide rails, jamming on one side, and abnormally increased frictional resistance of the water-stop rubber. Sudden collisions or foreign objects blocking the gate during the opening and closing process may cause serious accidents such as damage to the gate structure or even breakage of the wire rope.

[0027] Existing methods mostly rely on external limit switches or fixed time windows to capture samples, lacking robust sensing capabilities for the start-stop characteristics of the current itself. This results in the sample boundaries being greatly affected by human settings, and the static bias error caused by sensor zero drift is not systematically processed, making it difficult to guarantee the consistency of input data under long-term operation.

[0028] In feature extraction, simple effective value calculation or global spectrum analysis is often used without decoupling the three-phase current at the mechanism level. This leads to the numerical characteristics of uneven force on one side and the overall load increase being confused. Furthermore, the impact detection is easily overwhelmed by power frequency sinusoidal fluctuations and stable load changes, and there is a lack of adaptive suppression methods for non-stationary backgrounds.

[0029] In addition, existing statistical feature extraction methods mostly focus on the global mean, variance and other summary indicators of the entire sequence, ignoring the differences in the location of abnormal events at different stages of start-up, operation and stop. This causes early local weak anomalies to be smoothed and diluted during the long sequence averaging process, resulting in the loss of key temporal location clues.

[0030] Existing recognition models mostly use fixed-scale feature extractors or single-layer fully connected classifiers, which make it difficult to simultaneously capture two abnormal patterns with significant differences in time scale: short-term spike impacts and continuous jamming across multiple segments. Furthermore, in industrial scenarios where samples are scarce, relying solely on classification boundaries can easily lead to class overlap and confusion, and the robustness to the recognition of marginal categories with small samples is insufficient.

[0031] In view of the above, this disclosure provides a method for monitoring the operation status of a sluice gate, aiming to solve one or more problems in the prior art. The implementation details of the technical solution of the embodiments of this disclosure are described in detail below.

[0032] Figure 1 This schematic diagram illustrates a flowchart of a sluice gate operation status monitoring method according to an exemplary embodiment of this disclosure. Figure 1 As shown, the sluice gate operation status monitoring method includes steps S101 to S104: Step S101: When the opening and closing action of the sluice gate is detected, a single operation segment is extracted from the continuous monitoring data, and the original three-phase current sequence is constructed based on the single operation segment. Step S102: Decouple the original three-phase current sequence to obtain the impact-sensitive feature sequence, and construct a feature matrix based on the impact-sensitive feature sequence; the impact-sensitive feature sequence includes one or more of the following: instantaneous unbalance degree sequence, equivalent positive sequence current amplitude sequence, positive sequence current low-frequency trend sequence, and total impact intensity sequence; Step S103: Input the feature matrix into the trained state recognition model to perform multi-scale pseudo-temporal coding and category prototype constraint classification using the state recognition model, and obtain the predicted probability of each state category output by the state recognition model. Step S104: The state category corresponding to the maximum predicted probability is taken as the current sluice gate operating state.

[0033] In the technical solutions provided by some embodiments of this disclosure, on the one hand, an adaptive start-stop interception mechanism is used to extract single-run segments from continuous monitoring data to construct the original three-phase current sequence, providing a standardized three-phase synchronous input for subsequent analysis and avoiding the problem of inconsistent boundaries caused by different start-stop durations and inconsistencies in start and stop times. On the other hand, a physical mechanism-driven feature decoupling system is constructed to decouple the original three-phase current sequence to obtain one or more of the following impact-sensitive feature sequences: instantaneous imbalance sequence, equivalent positive sequence current amplitude sequence, positive sequence current low-frequency trend sequence, and total impact intensity sequence. The mixed current signal with power frequency as the carrier is decomposed into independent physical channels such as structural imbalance, load trend, and transient impact, extracting a more accurate feature matrix and improving the accuracy of state monitoring. Furthermore, a lightweight state recognition model jointly optimized by multi-scale pseudo-temporal convolution and category prototype comparison constraints is introduced, taking into account both local impacts and cross-segment continuous anomalies, and forcing similar samples to cluster compactly in the embedding space. Under small sample conditions, the clarity of the distinction between slight jamming and normal heavy load is improved, further enhancing the accuracy of state category recognition.

[0034] The following will describe in more detail each step of the sluice gate operation status monitoring method in this example embodiment, with reference to the accompanying drawings and embodiments.

[0035] In step S101, when the opening and closing action of the sluice gate is detected, a single operation segment is extracted from the continuous monitoring data, and the original three-phase current sequence is constructed based on the single operation segment.

[0036] Specifically, at the sluice gate operation site, the sluice gate operation monitoring system continuously collects the three-phase current of the opening and closing motor, including the first phase current, the second phase current, and the third phase current, which can also be referred to as phase A current, phase B current, and phase C current. When a new opening and closing action of the sluice gate is detected, the identification of the sluice gate's operating status is triggered.

[0037] Because the duration of different start-up and stop processes varies and the start and stop times are not completely consistent, directly inputting the entire original monitoring data into the subsequent identification process can easily lead to inconsistent boundaries.

[0038] Therefore, in step S101, a single running segment is first extracted from the continuous monitoring data using an adaptive start-stop extraction mechanism. In one embodiment of this disclosure, the extraction of a single running segment from the continuous monitoring data includes: Based on the continuous monitoring data, the three-phase currents at each sampling time (or sampling point) are extracted, and the value with the largest absolute value among the three-phase currents is used as the judgment value of the sampling time; the three-phase currents include the first phase current, the second phase current, and the third phase current; The sampling time when the judgment value first exceeds the start threshold is marked as the start time; Starting from the start time, the sampling time when the judgment value is continuously lower than the stop threshold for a preset duration is marked as the end time; Several sampling times are reserved before the start time and several sampling times are reserved after the end time to capture the single running segment.

[0039] Specifically, the continuously acquired three-phase current monitoring current is scanned point by point to monitor the maximum value of the absolute value of the three-phase current at each sampling time.

[0040] The moment when the maximum absolute value of the three-phase current first exceeds the starting threshold is recorded as the start time of a single operating segment. The starting threshold is denoted as... , This represents the current threshold used to determine when the sluice gate begins to operate. In a convenient implementation method... You can use 0.1 to 0.2 times the rated current.

[0041] The scan continues forward from the start time. When the maximum absolute value of the three-phase current continuously falls below the stop threshold for a preset duration, the corresponding time is recorded as the end time of a single running segment. The stop threshold is denoted as... , This represents the current threshold used to determine the end of the sluice gate's operation. In a convenient implementation method... It can be taken as 0.05 to 0.1 times the rated current; the continuous length is denoted as... , This indicates the continuous low current determination length, which can be from 100 sampling times to 300 sampling times.

[0042] In addition, to preserve the short-term impact information at the start and stop moments, a reserved sampling time is added before the start time, and a delayed sampling time is added after the end time, together forming a single running segment. The number of reserved sampling times is denoted as... The number of delayed sampling times is recorded as These two values ​​represent the lengths of the pre-start buffer and the post-end buffer, respectively, and can be set to 40 to 100 sampling times. The extracted single-run segment is saved as a three-phase synchronous current segment.

[0043] Based on the above method, the adaptive start-stop interception mechanism can extract a single running segment from continuous monitoring data, which can completely cover the entire process of "start-run-stop", making it convenient to analyze imbalance, overall load and collision impact within the same running cycle.

[0044] After capturing a single operational segment, an original three-phase current sequence is constructed based on that segment. Specifically, the instantaneous current value sequences of phase A, phase B, and phase C are aligned according to their sampling times to form the original three-phase current sequence. Among them, the first The instantaneous values ​​of phase A current, phase B current, and phase C current at each sampling time are denoted as follows: , and , Indicates the sampling index.

[0045] In one embodiment of this disclosure, the method further includes: Based on the average value of the three-phase current at all sampling times during the single operation segment, zero offset is performed on the single operation segment; and / or The single running segment is linearly resampled along the time axis, and the linearly resampled single running segment includes a preset number of sampling times.

[0046] Specifically, to avoid strong zero-drift interference in a single operating segment, zero-bias correction can be performed on the single operating segment. That is, zero-bias correction is performed on the single operating segment based on the average value of the three-phase current at all sampling times.

[0047] Specifically, the reserved sampling time before the start time is taken first, and the average value of phases A, B, and C within the reserved sampling time is calculated respectively. Then, the average value is subtracted point by point from the entire current segment of the corresponding phase. When the reserved sampling time is insufficient, the median of the first 1% of the length range of a single operation segment can be taken as the zero bias estimate.

[0048] In one embodiment, for example, during a single sluice gate operation, 50 sampling points are reserved before the start time of the captured single operation segment. Within these 50 reserved sampling points, the average value of the instantaneous A-phase current is calculated to be 0.02A (amperes). Then, 0.02A is subtracted from each sample value in the A-phase current sequence of this sample (including reserved points, data points throughout the start-run-stop process, and delayed points). If an original sample value is 10.02A, the corrected value is 10.00A. The processing method for phases B and C is the same, subtracting the average value within their respective reserved segments.

[0049] Based on the above method, the zero drift of the current sensor and the static bias of the acquisition circuit can be effectively eliminated, so that the signal returns to zero when no current passes through.

[0050] To avoid inconsistent lengths in single-run segments, the length of each segment can be standardized. This involves linearly resampling the single-run segment along the time axis, resulting in a pre-defined number of sampling moments within the resampled segment.

[0051] Specifically, N sampling times are preset as the total number of sampling times after the length is unified, for example, it can be unified to 2000 sampling times. Thus, the original three-phase current sequence... It refers to a three-phase synchronous input with a uniform length and a size of [missing information]. The sampling index n ranges from 1 to... .

[0052] If the original length of a single running segment is greater than Then, linear resampling and compression are performed according to the time axis to... Each sampling time; if the original length of a single running segment is less than [number] sampling times; Then, linear resampling along the time axis is performed to extend the process. Each sampling time.

[0053] In practical implementation, if the sampling time interval is denoted as , This represents the time interval between two adjacent sampling points, for example, it can be 0.0005 seconds. At that time, the unified single-run segment coverage duration is 1 second. If the original run duration is different from 1 second, the relative time order of the entire run will still be retained after linear resampling, without changing the sequential relationship of the run stages.

[0054] Based on the above method, the length of a single running segment can be standardized, which facilitates providing consistent input for subsequent mechanism decoupling and feature extraction.

[0055] It should be noted that zero-bias correction and length unification can be performed either one or both, and the order is not limited. For example, zero-bias correction can be performed first, and then the length of the single-run segment after zero-bias correction can be unified.

[0056] Figure 2 This diagram schematically illustrates a primitive three-phase current sequence in an exemplary embodiment of this disclosure. Figure 2 As shown, the original three-phase current sequence constructed after the above-described truncation, zero-bias correction, and length unification are illustrated. The figure plots the curves of the instantaneous values ​​of the currents in phases A, B, and C as a function of time. The horizontal axis represents time in seconds (s), and the vertical axis represents the instantaneous current value in amperes (A).

[0057] The graph clearly shows the current evolution during a complete opening and closing of the sluice gate: the current rises rapidly with a short-term surge during the start-up phase, the current remains at a high level and the three phases are basically symmetrical during the stable operation phase, and the current gradually decreases during the shutdown phase. The graph shows slight amplitude differences in the three-phase currents during the stable period, reflecting the potential three-phase imbalance that may exist in actual operation.

[0058] Based on the above, this disclosure proposes an adaptive start-stop interception mechanism based on the maximum absolute value of the three-phase current. Combined with dynamic zero-bias correction of the pre-start reserved segment and uniform-length resampling, this solves the problems of inconsistent sample boundaries and sensor static bias interference under variable-duration operation cycles, providing a standardized three-phase synchronous input for subsequent analysis.

[0059] In step S102, the original three-phase current sequence is decoupled to obtain the impact-sensitive feature sequence, and a feature matrix is ​​constructed based on the impact-sensitive feature sequence.

[0060] Specifically, during the operation of a sluice gate, the three-phase current simultaneously contains information about normal motor drive, asymmetry caused by uneven force distribution on the left and right sides, and short-term impact information caused by opening / closing collisions and sudden jamming. If only the effective value of the current is calculated or only a global spectrum analysis is performed, it is easy to confuse "overall load increase" with "unilateral structural abnormality," and it is also difficult to stably characterize the moments of opening / closing and sudden jamming. Therefore, this disclosure decomposes the original three-phase current sequence into multiple impact-sensitive feature sequences, and then further constructs a feature matrix.

[0061] The impact-sensitive characteristic sequence includes one or more of the following: instantaneous imbalance sequence, equivalent positive sequence current amplitude sequence, positive sequence current low-frequency trend sequence, and total impact intensity sequence.

[0062] In one embodiment of this disclosure, the specific process of decoupling to obtain the instantaneous imbalance sequence is as follows: Based on the instantaneous current value sequence of each phase in the original three-phase current sequence, the sliding effective value is calculated according to the sliding window to obtain the amplitude sequence of each phase; For each sampling time, the average value of the three-phase amplitude is calculated based on the amplitude sequence of each phase as the average value of the three-phase common load water. The standard deviation of the three-phase amplitude relative to the average value of the three-phase common load level is calculated as the three-phase common load level dispersion. Divide the three-phase common load level dispersion by the three-phase common load level average to obtain the instantaneous unbalance value corresponding to the sampling time; Traverse all sampling times in the original three-phase current sequence to obtain the instantaneous unbalance sequence composed of the instantaneous unbalance values; Specifically, unilateral jamming, guide rail wear, and mechanical connection misalignment during sluice gate operation typically manifest first as deviations between local amplitudes of the three phases, rather than a simultaneous rise of all three phases. In calculating the instantaneous unbalance sequence, this disclosure first calculates the three-phase amplitude sequence to obtain the amplitude sequence of each phase, and then calculates the instantaneous unbalance sequence based on the three-phase amplitude sequence, thus separating "structural imbalance" from "overall load increase."

[0063] First, the three-phase amplitude sequence is calculated based on the three-phase current instantaneous value sequence in a single operation segment. That is, based on the instantaneous current value sequence of each phase in the original three-phase current sequence, the sliding effective value is calculated according to the sliding window to obtain the amplitude sequence of each phase.

[0064] The sliding RMS window length is determined based on the power frequency and the sampling time interval. The power frequency is denoted as... , This represents the power frequency of the motor start / stop power supply, for example, 50Hz. The length of the sliding effective value window is denoted as... , This represents the number of sampling points used to estimate the local current amplitude, preferably the number of sampling points corresponding to one power frequency cycle. , This indicates a rounding operation. For example, when... Hz, seconds, .

[0065] The sliding effective value is calculated for the instantaneous current sequences of phase A, phase B, and phase C, respectively, to obtain the amplitude sequences of phase A, phase B, and phase C.

[0066] Taking the amplitude sequence of phase A as an example, the first... The amplitude of phase A at each sampling time is denoted as , is represented as: ; in, Indicates the first Local amplitude estimates of phase A at each sampling time; This represents the relative position index within the sliding window.

[0067] In practical implementation, the calculation methods for the B-phase amplitude sequence and the C-phase amplitude sequence are the same as those for the A-phase, respectively obtaining the first... B-phase amplitude at each sampling time and the C-phase amplitude at each sampling time .

[0068] In practical implementation, the starting position of the sequence cannot be directly calculated due to insufficient sliding window space. For each sampling point, boundary filling is performed using the first effective value replication method, ensuring that the A-phase amplitude sequence, B-phase amplitude sequence, and C-phase amplitude sequence are consistent with the original three-phase current sequence. Having the same length .

[0069] In one embodiment, for example, if the mean square of the current in phase A is 100, the mean square of the current in phase B is 121, and the mean square of the current in phase C is 81 within a local window of a certain operation sample, then the amplitudes of phase A, phase B, and phase C at the corresponding time are 10, 11, and 9, respectively. Based on this, the obtained three-phase amplitudes reflect the local load level, and no longer retain the rapid positive and negative oscillations of the power frequency sine wave.

[0070] Figure 3 This diagram schematically illustrates a three-phase amplitude sequence according to an exemplary embodiment of the present disclosure. Figure 3 The figure shows the three-phase amplitude sequence obtained after calculating the sliding effective value of the original three-phase current sequence. The horizontal axis represents time in seconds (s); the vertical axis represents the effective value of the current (i.e., the amplitude estimate) in amperes (A). The curves have filtered out the rapid positive and negative oscillations of the power frequency sine wave, retaining only the slowly varying envelope of the local load level. The amplitude variation trends of the three stages of start-up, steady-state, and shutdown can be seen from the figure. At the same time, the amplitude of phase B is slightly higher than that of phases A and C, reflecting the inconsistency of the three-phase load.

[0071] After obtaining the three-phase amplitude sequence, the instantaneous imbalance sequence is calculated. For each sampling time... Calculate the average value of phase A amplitude, phase B amplitude, and phase C amplitude as the average value of the three-phase common load water. , Indicates the first The average estimated value of the three-phase common load level at each sampling time can also be denoted as the three-phase average amplitude.

[0072] Furthermore, for each sampling time Calculate the amplitudes of phase A, phase B, and phase C relative to... The standard deviation is used to obtain the dispersion of the three-phase common load level. , Indicates the first The degree of inconsistency in the local amplitudes of the three phases at each sampling time can also be denoted as the three-phase dispersion.

[0073] The instantaneous imbalance sequence is then obtained by dividing the three-phase dispersion by the three-phase average amplitude. , is represented as: ; in, Indicates the first The instantaneous imbalance value at each sampling time; This represents a minimal constant to prevent the denominator from being zero; for example, it can be taken as... . The larger the value, the more inconsistent the three-phase amplitudes are at that moment, and the higher the probability of abnormal force on one side or structural tilt.

[0074] Furthermore, it is also possible to analyze the instantaneous imbalance sequence. Short-window smoothing is performed to suppress isolated spikes caused by jitter in individual sampling points. The short-window length is denoted as... , This represents the smoothing window length of the instantaneous imbalance sequence, which can be, for example, 3 to 5 sampling points.

[0075] In one embodiment, for example, if at a certain moment the amplitude of phase A is 10, the amplitude of phase B is 10.5, and the amplitude of phase C is 9.5, then the three-phase average amplitude is 10, the three-phase dispersion is approximately 0.408, and the instantaneous imbalance value is approximately 0.0408; if at another moment the amplitude of phase A is 12, the amplitude of phase B is 10, and the amplitude of phase C is 9.8, then the three-phase average amplitude is approximately 10.6, the three-phase dispersion is approximately 0.993, and the instantaneous imbalance value is approximately 0.0937. It can be seen that even if the overall current is at a high level in both cases, the instantaneous imbalance value will still be higher when the unilateral anomaly is more pronounced.

[0076] It should be noted that many early mechanical anomalies during sluice gate operation do not involve a "general increase in size," but rather a "first abnormality on one side, followed by three-phase imbalance." By directly quantifying the consistency of the local amplitudes of the three phases, the degree of structural imbalance can be stably reflected without introducing complex symmetrical component calculations. Based on this, under normal heavy-load conditions where the overall load increases synchronously, the instantaneous imbalance sequence will not be significantly elevated; however, under conditions such as unilateral jamming or guide rail wear, the instantaneous imbalance sequence will show a local increase earlier.

[0077] Figure 4 This illustration schematically depicts a sequence of instantaneous imbalances according to an exemplary embodiment of the present disclosure. Figure 4 As shown, the instantaneous unbalance sequence is calculated from the three-phase amplitude sequence and smoothed by a short window. The horizontal axis represents time in seconds (s); the vertical axis represents the instantaneous unbalance, a dimensionless ratio, with a larger value indicating greater inconsistency in the three-phase amplitudes. The figure shows that the unbalance remains at a low level during the stable operation phase, while localized increases occur near startup and shutdown, consistent with the physical phenomena of uneven force distribution or temporary off-center loading during startup and shutdown.

[0078] This disclosure can separate "structural imbalance" information from "overall load increase" through instantaneous imbalance degree, providing an independent sensitive feature for early identification of anomalies such as unilateral jamming and guide rail wear, and this indicator will not be significantly increased under normal heavy load conditions.

[0079] In one embodiment of this disclosure, the specific process of decoupling to obtain the equivalent positive sequence current amplitude sequence is as follows: For each sampling time Calculate the average values ​​of phase A, phase B, and phase C amplitudes to obtain the equivalent positive sequence current amplitude sequence. , Indicates the first The equivalent positive sequence current amplitude at each sampling time is used to characterize the strength variation of the three-phase driven load.

[0080] Since the average value of the three-phase amplitude is taken, the impact of local high or low values ​​of a single phase on the sequence is weakened, and the overall load change is preserved.

[0081] In one embodiment of this disclosure, the specific process of decoupling to obtain the low-frequency trend sequence of positive-sequence current is as follows: For the equivalent positive sequence current amplitude sequence Perform low-pass smoothing to obtain the low-frequency trend sequence of positive-sequence current. Positive sequence current low-frequency trend sequence It represents the overall load evolution trend throughout the opening and closing process, mainly reflecting the slow load change characteristics of the gate during the start-up, uniform speed operation and stop stages.

[0082] In one implementation, the equivalent positive sequence current amplitude sequence Execution length is The moving average, Indicates the length of the trend smoothing window, taking priority as... arrive For example, when hour, You can take 200 to 400.

[0083] In one embodiment of this disclosure, the specific process of decoupling to obtain the total impact intensity sequence is as follows: Based on the instantaneous current value sequence of each phase in the original three-phase current sequence, the instantaneous change intensity value of each sampling time compared with the adjacent sampling time is calculated to obtain the instantaneous change intensity sequence of each phase. The local baseline is calculated based on the instantaneous change intensity sequence of each phase, and the effective impact change value of each phase is obtained by subtracting the local baseline from the instantaneous change intensity value of each phase. The effective impact change value of each phase is fused into three phases to obtain the total impact intensity sequence.

[0084] Specifically, in order to extract short-term mechanical impact, the absolute value of the difference between adjacent sampling points is calculated for the instantaneous value sequences of phase A current, phase B current, and phase C current, respectively, to obtain the instantaneous change intensity sequences of phase A, phase B, and phase C.

[0085] In practical implementation, taking phase A as an example, the first... The instantaneous change intensity value of phase A at each sampling time is denoted as... The calculation method is as follows Furthermore, the calculation methods for phase B and phase C are the same, yielding the results for phase B and phase C respectively. Instantaneous change intensity value of phase B at each sampling time and the Instantaneous change intensity value of phase C at each sampling time .

[0086] Local baselines are calculated for the instantaneous intensity change sequences of phase A, phase B, and phase C, respectively. The effective impact change values ​​of phase A, phase B, and phase C are obtained by subtracting the local baselines from the current intensity change.

[0087] Taking phase A as an example, in practical implementation, the length can be taken as... The moving average is used as the local baseline, denoted as . Press again The effective impact change value of phase A is obtained by calculation, where, This indicates the operation of taking the maximum value. Furthermore, the calculation methods for phases B and C are the same, yielding the following results: and .

[0088] in, Indicates the first At each sampling moment, the intensity of the sharp current change caused by abnormal events such as short-term mechanical shocks, after removing normal power frequency fluctuations and routine fluctuations during stable operation, is a non-negative value. The larger the value, the more violent the instantaneous current change that occurred at that moment. Indicates the first At each sampling moment, the remaining sharp current change intensity caused by abnormal events such as short-term mechanical shocks is the current intensity after removing normal power frequency fluctuations and regular fluctuations during stable operation from the B-phase current. Indicates the first At each sampling moment, the intensity of sharp current changes caused by abnormal events such as short-term mechanical shocks is the remaining value in the C-phase current after removing normal power frequency fluctuations and regular fluctuations during stable operation.

[0089] It should be noted that when the sluice gate's opening and closing motor is operating normally, its current waveform is itself a sine wave. Even under steady-state conditions, there is a normal difference between adjacent sampling points due to variations in the power frequency (i.e., (It is not zero itself). Furthermore, the smooth change in load during start-up and shutdown will also produce a gradual change in current. If used directly... To characterize the impact, these normal, non-impact fluctuations constitute a significant amount of background noise. This is achieved by subtracting the average change intensity within a local sliding window (i.e., the local baseline). This can adaptively eliminate the background effects of normal power frequency fluctuations and stable load changes. The final result is... It only responds to sharp changes that significantly exceed the level of normal local fluctuations. Based on this, the extraction of impact features has background adaptive capabilities, and can more cleanly and stably highlight current transients truly caused by mechanical impacts, such as opening and closing collisions and sudden jamming.

[0090] The effective impact variation values ​​of phase A, phase B, and phase C are fused to obtain the total impact intensity sequence. , is represented as: ; in, Indicates the first The total impact intensity value at each sampling time; This indicates the operation of finding the minimum value; This represents the three-phase synchronization enhancement coefficient, used to amplify the mechanical impact response corresponding to the three-phase synchronization mutation. For example, it can be taken as 0.3 to 1.0. The term is used to reflect the overall mutation intensity, and the minimum term is used to ensure that the total impact intensity value is further increased only when all three phases show significant changes.

[0091] It should be noted that mechanical collisions or sudden jamming are usually synchronous physical events affecting all three phases. Therefore, the minimum value term will only be large when all three phase currents simultaneously exhibit significant impact changes, thus affecting the coefficient. Significantly increase total impact strength Conversely, if only a single phase experiences an isolated spike due to sensor noise or other reasons (in which case the minimum value will be very small), the contribution of this term to the total impact intensity will be very small, thus effectively suppressing misjudgments caused by non-mechanical impacts.

[0092] In addition, the total impact intensity sequence can also be analyzed. Perform short-window smoothing, with the smoothing window length denoted as... , This represents the smoothing window length of the total impact intensity sequence, which can be, for example, 3 to 5 sampling points.

[0093] Based on this, the original The sequence may contain isolated spikes at the level of individual sampling points due to sampling noise or computational processes. By performing smoothing (e.g., moving average) over a short window (e.g., 3 to 5 sampling points), these high-frequency spikes can be filtered out without significantly reducing the accuracy of impact event timing. This allows the true impact response to form a more continuous and prominent local peak envelope. Based on this, the position and intensity of the peak in the sequence are more stable and reliable, less prone to fluctuations due to jitter at individual sampling points, and easier for subsequent feature extraction (e.g., counting the number of local peaks).

[0094] In one embodiment, if the effective impact change values ​​of phase A, phase B, and phase C at a certain moment are 1.2, 1.1, and 1.3 respectively, then... The total impact strength value is If at another moment only phase A has single-point noise, the corresponding effective impact change values ​​for phase A, phase B, and phase C are 1.2, 0.1, and 0.1, respectively. Then the total impact intensity value is... Therefore, it can be seen that the total impact intensity value corresponding to the three-phase synchronous sudden change is significantly higher, which is more consistent with the actual mechanism of mechanical collision and jamming.

[0095] It should be noted that abnormal sluice gate operation can manifest as either a slow increase in load or a short-term synchronous impact, with different time scales and physical meanings. This step describes "commonly driven load" and "synchronous impact event" separately, rather than using a single current curve to represent both meanings simultaneously. Based on this, the positive sequence current low-frequency trend sequence is more suitable for characterizing the overall operating resistance change, while the total impact intensity sequence is more suitable for characterizing collisions, sudden jamming, and hard impacts during opening and closing. The combination of the two can simultaneously cover both slow and fast anomalies.

[0096] Figure 5 This illustration schematically depicts an equivalent positive-sequence current amplitude and low-frequency trend sequence in an exemplary embodiment of this disclosure. For example... Figure 5 As shown, the graph contains two curves: the solid blue line represents the equivalent positive-sequence current amplitude sequence, which approximates the strength of the load driven by the three phases by the average of the three-phase amplitudes; the dashed black line represents the low-frequency trend sequence of the positive-sequence current, which extracts the overall evolution trend of the load through a long-window sliding average. The horizontal axis represents time in seconds (s); the vertical axis represents current amplitude in amperes (A). As can be seen from the graph, the equivalent positive-sequence amplitude curve fully reflects the load changes throughout the entire start-up and shutdown process, while the low-frequency trend curve further smooths out local fluctuations, highlighting the slow-changing characteristics of load increase during startup, stable operation, and load decrease during shutdown.

[0097] Figure 6 This schematic diagram illustrates a total impact intensity sequence in an exemplary embodiment of the present disclosure. Figure 6 The figure shows the total impact intensity sequence after three-phase synchronous fusion and short-window smoothing. The horizontal axis represents time in seconds (s); the vertical axis represents the total impact intensity value, which reflects the degree of sharp current mutation remaining after removing normal power frequency fluctuations and stable load changes, and can be understood as the impact intensity within a unit sampling interval. A significant peak appears at start-up and shutdown times, accurately capturing the three-phase synchronous current mutation caused by mechanical collisions or sudden jamming; the intensity is close to zero during stable operation, indicating that there are no abnormal impacts during normal drive. This figure demonstrates that the present invention, through differential baseline subtraction and three-phase synchronous enhancement mechanisms, can adapt to background fluctuations and stably highlight current transients truly caused by mechanical impacts, providing a high signal-to-noise ratio feature sequence for the identification of collision anomalies and severe jamming.

[0098] Based on the above, this disclosure constructs a physical mechanism-driven feature decoupling system. It separates single-sided jamming from overall heavy load by instantaneous imbalance, and extracts pure impact intensity by differential baseline subtraction and three-phase synchronous minimum value enhancement. It clearly decomposes the mixed current signal with power frequency as the carrier into three independent physical channels: structural imbalance, load trend and transient impact.

[0099] After obtaining the sensitive feature sequence through decoupling, a feature matrix is ​​constructed based on the shock sensitive feature sequence.

[0100] After the aforementioned processing, the instantaneous imbalance sequence has been obtained. Equivalent positive sequence current amplitude sequence Total impact intensity sequence and positive sequence current low frequency trend sequence To transform the sensitive feature sequences into inputs that can be directly used by subsequent recognition models, this disclosure simultaneously extracts global statistical features and segmented positional features to construct a feature matrix. .

[0101] For the instantaneous imbalance sequence Equivalent positive sequence current amplitude sequence Total impact intensity sequence and positive sequence current low frequency trend sequence Calculate six global statistics: mean, standard deviation, maximum value, root mean square value, 95th percentile value, and peak factor.

[0102] Among them, the mean is used to describe the overall level of the sequence; the standard deviation is used to describe the degree of volatility; the maximum value is used to describe the strongest anomaly amplitude; the root mean square value is used to describe the energy level; the 95th percentile value is used to describe the stable upper bound of the high response interval; and the peak factor is defined as the maximum value divided by the sum of the root mean square value and the minimum constant, used to describe the prominence of extreme peaks relative to the normal energy. Based on this, each sequence contributes 6 dimensions of global statistical features.

[0103] Divide each sequence into equal parts along the time axis. segment, number of segments This indicates the number of segments for retaining location information during the runtime phase; for example, 8 could be used. If... Cannot be If the result is divisible, the extra sampling points will be evenly distributed to the first few segments to ensure complete coverage of all sampling points.

[0104] Furthermore, the local energy percentage is calculated within each segment. The local energy percentage represents the proportion of the sum of squares of the samples in that segment to the sum of squares of the entire sequence, reflecting whether the anomalous response is more concentrated in the initiation, middle, or termination phases. For the first... The segment, the local energy percentage is denoted as , This represents a segment index, with values ​​ranging from 1 to... .

[0105] Furthermore, the number of local peaks within each segment is calculated. The number of local peaks is denoted as... , Indicates the first The number of significant peaks within a segment. In specific implementation, if a sampling point simultaneously satisfies "greater than the previous sampling point", "greater than or equal to the next sampling point", and "greater than the current segment mean plus..." If the standard deviation of the current segment is multiplied by 1, then the sampling point is counted as a local peak. This represents the peak determination coefficient, which can be 1.5 to 2.5, and in a more convenient implementation, it can be 2.

[0106] In one embodiment, as an example, suppose there is an instantaneous imbalance sequence for a certain run. Divided into equal parts Now we examine the calculation of the number of local peaks in segment 3. Let the mean of all sampling points in segment 3 be... The standard deviation is Peak determination coefficient The threshold condition for determining whether a sampling point is a local peak is: the value of that point... Within this segment, if the value of a certain sampling point is... And it satisfies the condition that it is greater than its previous sampling point (e.g. ) and greater than or equal to the next sampling point (e.g. If a point is found that satisfies all the above conditions, then that point is counted as one local peak. Therefore, the total number of sampling points within that segment that satisfy all the above conditions is the number of local peaks in that segment. .

[0107] The 6-dimensional global statistical features, 8-dimensional local energy proportion features, and 8-dimensional local peak count features of each sequence are concatenated in a fixed order to obtain the 22-dimensional feature row vector corresponding to the sequence.

[0108] The 22-dimensional feature row vectors are arranged in the following order: the first 6 dimensions are global statistical features, the 7th to 14th dimensions are 8 local energy proportion features, and the 15th to 22nd dimensions are 8 local peak count features.

[0109] According to the "instantaneous imbalance sequence" Equivalent positive sequence current amplitude sequence Total impact intensity sequence Positive sequence current low-frequency trend sequence The four 22-dimensional feature row vectors are stacked row by row in the order of "" to form a feature matrix. Feature matrix The multi-domain feature representation of a single execution segment is of fixed length and size. .

[0110] Furthermore, regarding the characteristic matrix Perform dimension-by-dimensional standardization. Specifically, compute the feature matrix on the training set. The mean and standard deviation of each feature column are calculated and saved as standardized parameters. During both the training and online recognition phases, the same set of standardized parameters is used to subtract the mean and divide by the standard deviation for the corresponding dimensions. This effectively eliminates the impact of differences in the dimensions of different statistical measures on the subsequent training of the recognition model.

[0111] In one embodiment, as an example, when At that time, each sequence outputs 22-dimensional features, and the four sequences yield a total of 22-dimensional features. Feature matrix For example, the total impact intensity sequence If the number of local peaks is high in segments 1 and 8, it indicates that the anomaly is more likely to occur at the start and stop moments; if the positive sequence current has a low-frequency trend sequence... If the local energy proportion is relatively high in several intermediate segments, it indicates that the overall resistance level in the middle section of operation remains high.

[0112] It should be noted that sluice gate anomalies are not only related to their magnitude but also closely related to the operational stage in which they occur. For example, the hard impact during startup and the persistent jamming during mid-operation have different mechanisms. This step does not simply calculate the mean or peak value of the entire sequence, but rather retains both "overall strength" and "location of occurrence" information. Based on this, without significantly increasing the feature dimension, the feature matrix... Sufficient temporal structure information and stage location information are still retained.

[0113] Based on the above, this disclosure adopts a fixed-length feature matrix construction method that preserves temporal position information. The decoupled sequence is segmented along the time axis to statistically analyze the energy ratio and the number of peaks. Without relying on recurrent neural networks, a compact statistical feature matrix simultaneously encodes the magnitude of abnormal amplitudes and their occurrence positions in the start-up and shutdown phases.

[0114] Step S103: Input the feature matrix into the trained state recognition model to perform multi-scale pseudo-temporal coding and category prototype constraint classification using the state recognition model, and obtain the predicted probability of each state category output by the state recognition model.

[0115] Specifically, the feature matrix Internally, it still contains a two-layer structure: "sources of different mechanistic features" and "distribution locations of the same mechanistic feature at different operational stages." If a single-layer fully connected classifier is used directly, it's easy to overlook the connections between adjacent segments and it's difficult to stably distinguish between classes with similar boundaries. Therefore, this disclosure uses a feature matrix... As a multi-channel pseudo-temporal input, multi-scale pseudo-temporal convolution is used to extract cross-segment associations, and then combined with category prototype constraints to achieve lightweight recognition.

[0116] In one embodiment of this disclosure, the method further includes: pre-training the state recognition model; the pre-training of the state recognition model includes: For a single running sample in the running sample set, the impact-sensitive feature sequence of the running sample is extracted to construct a feature matrix, and an initial input feature map is constructed based on the feature matrix; The input feature map is encoded using a convolutional network to obtain an encoded output feature map; the convolutional network includes multiple convolutional blocks of different scales, and a gating network is used to adaptively generate the scale weights of each convolutional block for feature fusion. The predicted probability of each state category is calculated based on the encoded output feature map, and the total loss value is calculated based on the predicted probability. The model parameters of the convolutional network are adjusted using the total loss value until training is complete and the state recognition model is obtained.

[0117] Specifically, the proposed identification model is trained, validated, and deployed using labeled historical operating data, enabling the model to perform real-time status identification at the actual sluice gate operation site.

[0118] First, a set of operational samples is constructed, comprising multiple operational samples, including operational data and labeled real-world category statuses. The collected operational samples are then labeled; labels can be determined by combining maintenance records, video verification, manual inspection results, and operational logs.

[0119] In one implementation, the total number of state categories Four categories are defined as follows: Normal opening and closing: The opening and closing process is smooth, without obvious jamming, collision, or abnormal stoppage; Slight jamming: The opening and closing process can be completed, but there is local obstruction, increased local friction, or short-term aggravation of imbalance; Severe jamming: The resistance during the opening and closing process increases significantly and lasts for a long time, which may be accompanied by obvious off-center load or multiple abnormal stoppages; Abnormal collision: There is an obvious synchronous impact peak at the moment of start-up, the moment of stop, or during operation, corresponding to hard collision, rebound, or sudden impact event.

[0120] All labeled samples are divided into training, validation, and test sets. The training set is used for parameter learning, the validation set is used for model selection, and the test set is used for final performance evaluation.

[0121] For each running sample in the training set, steps S101 and S102 are executed sequentially to construct the feature matrix. Then, the feature matrix... Before being fed into the recognition model, the structure needs to be clearly defined and rearranged to construct the initial input feature map to adapt to the subsequent convolutional network.

[0122] Specifically, the feature matrix The four rows serve as four input channels, corresponding sequentially to the instantaneous imbalance sequence. Equivalent positive sequence current amplitude sequence Total impact intensity sequence and positive sequence current low frequency trend sequence 22-dimensional statistical characteristics.

[0123] Furthermore, the feature matrix The 22 columns are used as pseudo-temporal positions, where columns 1 to 6 correspond to the global statistical feature region, columns 7 to 14 correspond to 8 segments of local energy proportion feature region, and columns 15 to 22 correspond to 8 segments of local peak count feature region, thus obtaining the initial input feature map. Based on this, adjacent positions along the column direction possess clear statistical or stage-wise adjacency relationships. Initial input feature map This represents the initial feature map entering the multi-scale pseudo-temporal convolutional coding network, with a size of [size missing]. .

[0124] Different anomalies in the feature matrix The distribution scales differ. Initial hard impacts are typically concentrated in a few adjacent columns, while persistent stuttering often occurs across multiple adjacent segments. Therefore, this disclosure sets multiple convolutional scales simultaneously in each convolutional block and uses gating weights to adaptively fuse the multi-scale responses.

[0125] Specifically, the first The input feature maps of each convolutional block are denoted as follows: ,in, This represents the convolution block index.

[0126] In one implementation, the total number of convolutional blocks can be 3. The first convolutional block has 4 input channels and 16 output channels; the second and third convolutional blocks each have 16 input channels and 16 output channels.

[0127] For the input feature map Perform global average pooling along the pseudo-timing direction to obtain the channel description vector. Channel description vector It represents the overall response intensity of the current sample on each input channel, used to determine whether the current sample is more inclined to "local spike-type anomaly" or "cross-segment persistent anomaly".

[0128] Furthermore, the channel description vector Inputting the gating network yields three scale weights.

[0129] In practical implementation, the gated network can consist of two fully connected layers. The first layer maps the channel description vectors to the hidden dimensions, which can be, for example, 8, and uses a modified linear unit activation function. The second layer outputs three real numbers, which are then normalized by Softmax to obtain three scale weights. , and These correspond to branches with kernel lengths of 3, 5, and 7, respectively, and the sum of the weights of the three scales is 1.

[0130] Input feature map Simultaneously, three one-dimensional convolutional branches are fed in, and features are extracted along the pseudo-temporal direction using convolutional kernels with lengths of 3, 5, and 7, respectively, resulting in the output feature maps of the three branches. , and The three branches have the same number of output channels, and the pseudo-timing length is kept at 22 through symmetrical padding.

[0131] It should be noted that branches with a kernel length of 3 focus more on local mutations, branches with a kernel length of 5 take into account both local and mid-range relationships, and branches with a kernel length of 7 focus more on continuous changes across multiple segments.

[0132] The feature maps output from the three branches are weighted and fused according to their scale weights to obtain a fused feature map. , represented as ;in, Indicates the first The result of multi-scale fusion of convolutional blocks.

[0133] Furthermore, the fused feature map Batch normalization, activation of linear units with leakage correction, and so on are performed sequentially. Convolution is used to obtain the main branch output.

[0134] Among them, batch normalization is used to stabilize the training process, and leakage-corrected linear unit activation is used to enhance nonlinear representation. Convolution is used to perform channel mixing without changing the pseudo-timing length.

[0135] Input feature map By adding element-wise the residual branch to the main branch output, we obtain the first... Output feature map of each convolutional block When the number of input channels differs from the number of output channels, for example, when the first convolutional block changes from 4 channels to 16 channels, an additional channel is added to the residual branch. Convolution performs channel matching; when the number of input channels is the same as the number of output channels, for example, when the second and third convolutional blocks are both 16 channels, the residual branch directly passes the input feature map.

[0136] By concatenating the first, second, and third convolutional blocks sequentially, the encoded output feature map is obtained. Encode output feature map This represents the high-level state representation after multi-scale pseudo-temporal convolutional encoding, with a size of [size missing]. .

[0137] In one embodiment, if the total impact intensity sequence of a sample to be identified The relevant features show significantly enhanced responses in segments 1 and 8, while the responses in the middle segments are weaker. Therefore, the gating network is more likely to provide higher values. This gives greater weight to branches with a kernel length of 3, thus highlighting local impact patterns; if the positive sequence current low-frequency trend sequence If the gating network increases continuously across multiple consecutive segments, it is more likely to improve. This gives greater weight to branches with a kernel length of 7, thereby enhancing the ability to model persistent sluggish patterns.

[0138] It should be noted that sluice gate malfunctions exhibit two typical time scales: "local spike-type" and "spanning continuous type." This step does not fixate on a single convolutional scale, but rather adaptively fuses multiple convolutional scales based on the overall response of the current sample. Based on this, with a relatively small model size, it can remain sensitive to short-term shocks while also responding to sustained increases in resistance, avoiding missed detections caused by a single convolutional scale.

[0139] After multi-scale pseudo-temporal convolutional encoding, the encoded output feature map It already possesses strong discriminative capabilities, so the predicted probability of each state category can be calculated based on the encoded output feature map, and the total loss value can be calculated by comparing it with the labeled real results, so as to modify the network model parameters.

[0140] However, there may still be cases where the boundaries are close between different categories. In particular, slight jamming and normal heavy-load conditions are easily confused. Therefore, in this disclosure, in addition to the prototype contrast constraint loss term, a category-weighted cross-entropy loss term can be added. Through category prototype constraints, samples of the same type are more compact in the feature space, and samples of different types are more separated.

[0141] In one embodiment of this disclosure, the total loss value includes a class-weighted cross-entropy loss term and a prototype contrast constraint loss term. The step of calculating the predicted probability of each state class based on the encoded output feature map and calculating the total loss value based on the predicted probability includes: Global average pooling is performed on the encoded output feature map along the pseudo-temporal direction to obtain the sample embedding vector; Calculate the similarity between the sample embedding vector and the category prototype vector corresponding to each state category to obtain the predicted probability of each state category; The category-weighted cross-entropy loss term is calculated based on the predicted probability of each state category; and the prototype comparison constraint loss term is calculated based on the predicted state category corresponding to the maximum predicted probability and the actual state category of the running sample.

[0142] Specifically, the encoded output feature map Perform global average pooling along the pseudo-temporal direction to obtain the sample embedding vector. Sample embedding vector This represents the global state representation of a single execution segment, with a dimension equal to the encoded output feature map. The number of channels is 16 in one implementation.

[0143] Let the total number of state categories be denoted as , This indicates the number of sluice gate status categories that need to be identified. In one implementation, Four categories can be selected: normal opening and closing, slight jamming, severe jamming, and collision abnormality.

[0144] Furthermore, maintain a category prototype vector for each category. , This represents the category index, with values ​​ranging from 1 to... Category prototype vector Indicates the first The center position of the class sample in the embedding space, and its dimension relative to the sample embedding vector. same.

[0145] Further, the sample embedding vector is calculated. With prototype vectors of each category The cosine similarity between them yields the first... Class similarity , represented as ;in, Indicates the current sample and the first The degree of similarity between class prototypes; the higher the value, the closer to the center of the class.

[0146] Temperature scaling and Softmax normalization are applied to the similarity scores for each category to obtain the predicted probability that the current sample belongs to each category. The temperature coefficient is denoted as... , This represents a hyperparameter used to adjust the sharpness of the probability distribution, and can be, for example, taken as 0.1 to 0.3.

[0147] In practice, the category with the highest predicted probability is used as the identification result for the current sample.

[0148] During training, class-weighted cross-entropy loss and prototype contrast constraint are used for joint optimization. The class weight is denoted as , Indicates the first The amplification factor of class samples in the loss. To alleviate class imbalance, the first... The class weights can be calculated by dividing the number of samples in the largest class in the training set by the number of samples in the first class. The number of samples in each class is obtained. The total loss is denoted as... , represented as ;in, This represents the category-weighted cross-entropy loss; This represents the prototype comparison constraint weight, which can be from 0.05 to 0.2. This indicates the prototype comparison constraint.

[0149] In an easily implementable manner, the prototype is compared with the constraint terms. Defined as: ; in, This represents the true class index of the current sample; This indicates the similarity between the current sample and the true category prototype; This indicates an upper bound on the non-true category similarity, for example, it can be between 0 and 0.2. Based on this, samples will be pulled towards the true category prototype, while being pushed away from non-true category prototypes that are too close to them.

[0150] The initial learning rate for training can be chosen. The batch size can be 16 to 32, and the total number of training cycles can be 80 to 150.

[0151] After each training cycle, the F1 score for each category is calculated using the validation set, and the arithmetic mean of the F1 scores for each category is obtained to obtain the macro-average F1 score. The recognition model parameters, standardized parameters, and category prototype vectors for each category that achieve the highest macro-average F1 score are saved as the final deployment version.

[0152] After training is complete, save the following deployment parameters: Startup threshold Stop threshold Unified number of sampling points Length of sliding effective value window Trend smoothing window length Standardized parameters, multi-scale pseudo-temporal convolutional coding network parameters, and class prototype vectors for each category. This ensures that the training phase and the online recognition phase are completely consistent.

[0153] Among them, the category prototype vector It can be updated using an exponential moving average method.

[0154] Specifically, if the first in the current training batch The mean of the class sample embedding vectors is denoted as Then use In this way, update the first The class prototype vector is then normalized. This represents the prototype update momentum coefficient, which can be set to 0.99.

[0155] In practical implementation, if the current batch does not contain the first... For class samples, the first one is kept. The class prototype vector remains unchanged.

[0156] In one embodiment, as an example, if a test sample is in an instantaneous imbalance sequence and positive sequence current low frequency trend sequence All of these showed a persistently high level, while the total impact intensity sequence If it is not prominent, then the sample embedding vector of that sample is... The category prototype vector is more likely to be closer to the severely constrained category. If the total impact intensity sequence of a test sample in the starting and stopping phases If the value rises significantly while the middle section remains stable, then the sample embedding vector of that sample is... The category prototype vector is more likely to be closer to the collision anomaly category. .

[0157] It should be noted that the number of samples for sluice gate operation status is usually limited, and there are certain similarities in local statistics between slight jamming and normal heavy-load conditions. This step combines lightweight convolutional coding with class prototype constraints, rather than relying solely on simple parameterized classification boundaries. Based on this, similar samples are more concentrated in the embedding space, dissimilar samples are more separated, and classification stability is higher under small sample conditions.

[0158] After training the state recognition model and deploying it on-site, the feature matrix obtained in step S102 is... Input the state recognition model to obtain the predicted probability of each state category.

[0159] Step S104: The state category corresponding to the maximum predicted probability is taken as the current sluice gate operating state.

[0160] Based on the above, this disclosure introduces a lightweight recognition architecture jointly optimized by multi-scale pseudo-temporal convolution and class prototype contrast constraint. By adaptively fusing convolutional features from different receptive fields through a gating network, it can take into account both local impact and cross-segment continuous anomalies. Furthermore, it utilizes prototype loss to force similar samples to cluster compactly in the embedding space, thereby improving the clarity of the distinction between slight jamming and normal heavy loading under small sample conditions.

[0161] In one embodiment of this disclosure, an alarm can also be triggered if the maximum predicted probability is higher than an alarm threshold. The alarm threshold is denoted as... , This represents the probability threshold for triggering an abnormal alarm, and can be set to, for example, 0.7 to 0.9.

[0162] In one implementation, for example, if the current single-run segment is identified as slightly stuck, and the predicted probability is higher than... If the current single running segment is identified as severely stuck or experiencing collision anomalies, and the predicted probability is higher than 100%, the system will output a "minor stuttering warning"; if the current single running segment is identified as severely stuck or experiencing collision anomalies, and the predicted probability is higher than 100%, the system will output a "minor stuttering warning"; If this occurs, the system will output an alarm at the "Immediate Repair" level.

[0163] In addition, the identification results of multiple consecutive opening and closing actions can be saved in chronological order. If the same sluice gate is identified as slightly jammed multiple times within a preset time window, a trend maintenance reminder can be triggered; if collision anomalies or severe jamming occur multiple times in a short period of time, a shutdown inspection can be triggered directly.

[0164] Based on this, the system can not only identify single operational anomalies, but also discover the evolution trend of anomalies by combining continuous operational history.

[0165] In summary, this disclosure decomposes the original three-phase current sequence into an instantaneous imbalance sequence, an equivalent positive-sequence current amplitude sequence, a total impact intensity sequence, and a positive-sequence current low-frequency trend sequence. Then, by combining segmented statistical features and a multi-scale prototype comparison identification model, the state is determined. This allows for the separate expression of "structural imbalance," "overall load increase," and "short-term impact event," thereby improving the accuracy of state category identification.

[0166] Figure 7 This schematic diagram illustrates the composition of a sluice gate operation status monitoring device according to an exemplary embodiment of the present disclosure, such as... Figure 7 As shown, the sluice gate operation status monitoring device 700 may include a detection module 701, a decoupling module 702, a prediction module 703, and an output module 704. Wherein: The detection module 701 is used to extract a single operation segment from continuous monitoring data when the opening and closing action of the sluice gate is detected, and to construct an original three-phase current sequence based on the single operation segment. The decoupling module 702 is used to decouple the original three-phase current sequence to obtain the impact-sensitive feature sequence, and to construct a feature matrix based on the impact-sensitive feature sequence; the impact-sensitive feature sequence includes one or more of the following: instantaneous unbalance sequence, equivalent positive sequence current amplitude sequence, positive sequence current low-frequency trend sequence, and total impact intensity sequence; The prediction module 703 is used to input the feature matrix into the trained state recognition model, so as to use the state recognition model to perform multi-scale pseudo-temporal coding and category prototype constraint classification, and obtain the predicted probability of each state category output by the state recognition model. The output module 704 is used to take the state category corresponding to the maximum predicted probability as the current sluice gate operating state.

[0167] According to an exemplary embodiment of this disclosure, the detection module 701 is further configured to extract the three-phase current at each sampling time based on the continuous monitoring data, and take the value with the largest absolute value among the three-phase currents as the judgment value of the sampling time; the three-phase current includes the first phase current, the second phase current, and the third phase current; the sampling time when the judgment value first exceeds the start threshold is marked as the start time; after the start time, the sampling time when the judgment value is continuously lower than the stop threshold for a preset duration is marked as the end time; several reserved sampling times are retained before the start time, and several delayed sampling times are retained after the end time, so as to capture the single running segment.

[0168] According to an exemplary embodiment of this disclosure, the detection module 701 is further configured to perform zero offset on the single operating segment based on the average value of the three-phase current at all sampling times in the single operating segment; and / or linearly resample the single operating segment according to the time axis, wherein the linearly resampled single operating segment includes a preset number of sampling times.

[0169] According to an exemplary embodiment of this disclosure, when the impact-sensitive feature sequence includes an instantaneous imbalance sequence, the decoupling module 702 is further configured to calculate a sliding effective value based on the instantaneous current value sequence of each phase in the original three-phase current sequence according to a sliding window, to obtain the amplitude sequence of each phase; for each sampling time, calculate the average value of the three-phase amplitudes based on the amplitude sequence of each phase as the three-phase common load level average value; calculate the standard deviation of the three-phase amplitudes relative to the three-phase common load level average value as the three-phase common load level dispersion; divide the three-phase common load level dispersion by the three-phase common load level average value to obtain the instantaneous imbalance value corresponding to the sampling time; and traverse all sampling times in the original three-phase current sequence to obtain the instantaneous imbalance sequence composed of the instantaneous imbalance values.

[0170] According to an exemplary embodiment of this disclosure, when the impact-sensitive feature sequence includes a total impact intensity sequence, the decoupling module 702 is further configured to calculate the instantaneous change intensity value of each sampling time relative to the adjacent sampling time based on the instantaneous current value sequence of each phase in the original three-phase current sequence, so as to obtain the instantaneous change intensity sequence of each phase; calculate a local baseline based on the instantaneous change intensity sequence of each phase, and subtract the local baseline from the instantaneous change intensity value of each phase to obtain the effective impact change value of each phase; and perform three-phase fusion of the effective impact change value of each phase to obtain the total impact intensity sequence.

[0171] According to an exemplary embodiment of this disclosure, the sluice gate operation status monitoring device 700 further includes a model training module for pre-training the state recognition model, comprising: extracting the impact-sensitive feature sequence of an operation sample from a set of operation samples to construct a feature matrix, and constructing an initial input feature map based on the feature matrix; encoding the input feature map using a convolutional network to obtain an encoded output feature map; the convolutional network comprising multiple convolutional blocks of different scales, and using a gating network to adaptively generate the scale weights of each convolutional block for feature fusion; calculating the predicted probability of each state category based on the encoded output feature map, and calculating the total loss value based on the predicted probability; and adjusting the model parameters of the convolutional network using the total loss value until training is completed to obtain the state recognition model.

[0172] According to an exemplary embodiment of this disclosure, the total loss value includes a category-weighted cross-entropy loss term and a prototype contrast constraint loss term. The model training module is further configured to perform global average pooling on the encoded output feature map along the pseudo-temporal direction to obtain a sample embedding vector; calculate the similarity between the sample embedding vector and the category prototype vector corresponding to each state category to obtain the predicted probability of each state category; calculate the category-weighted cross-entropy loss term based on the predicted probability of each state category; and calculate the prototype contrast constraint loss term based on the predicted state category corresponding to the maximum predicted probability and the true state category of the running sample.

[0173] The specific details of each module in the aforementioned sluice gate operation status monitoring device 700 have been described in detail in the corresponding sluice gate operation status monitoring method, so they will not be repeated here.

[0174] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0175] In exemplary embodiments of this disclosure, a storage medium capable of implementing the above-described methods is also provided. It may be a portable compact disc read-only memory (CD-ROM) and include program code, and can run on a terminal device, such as a mobile phone. However, the program product of this disclosure is not limited thereto. In this document, a readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0176] In an exemplary embodiment of this disclosure, an electronic device capable of implementing the above-described method is also provided. Figure 8 The schematic diagram illustrates the structure of a computer system of an electronic device according to an exemplary embodiment of the present disclosure.

[0177] It should be noted that, Figure 8 The computer system 800 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0178] like Figure 8As shown, the computer system 800 includes a Central Processing Unit (CPU) 801, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 802 or programs loaded from storage section 808 into Random Access Memory (RAM) 803. The RAM 803 also stores various programs and data required for system operation. The CPU 801, ROM 802, and RAM 803 are interconnected via a bus 804. An Input / Output (I / O) interface 805 is also connected to the bus 804.

[0179] The following components are connected to I / O interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to I / O interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 810 as needed so that computer programs read from it can be installed into storage section 808 as needed.

[0180] In particular, according to embodiments of this disclosure, the processes described below with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by central processing unit (CPU) 801, it performs various functions defined in the system of this disclosure.

[0181] It should be noted that the computer-readable medium shown in the embodiments of this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0182] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0183] The units described in the embodiments of this disclosure can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the unit itself.

[0184] In another aspect, this disclosure also provides a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

[0185] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0186] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0187] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein.

[0188] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for monitoring the operational status of a sluice gate, characterized in that, include: When the opening and closing action of the sluice gate is detected, a single operation segment is extracted from the continuous monitoring data, and the original three-phase current sequence is constructed based on the single operation segment; The original three-phase current sequence is decoupled to obtain the impact-sensitive characteristic sequence, and a feature matrix is ​​constructed based on the impact-sensitive characteristic sequence; the impact-sensitive characteristic sequence includes one or more of the following: instantaneous unbalance degree sequence, equivalent positive sequence current amplitude sequence, positive sequence current low-frequency trend sequence, and total impact intensity sequence; The feature matrix is ​​input into the trained state recognition model to perform multi-scale pseudo-temporal coding and category prototype constraint classification, thereby obtaining the predicted probability of each state category output by the state recognition model. The state category corresponding to the maximum predicted probability is taken as the current sluice gate operating state.

2. The sluice gate operation status monitoring method according to claim 1, characterized in that, The step of extracting a single running segment from continuous monitoring data includes: Based on the continuous monitoring data, the three-phase currents at each sampling time are extracted, and the value with the largest absolute value among the three-phase currents is used as the judgment value of the sampling time; the three-phase currents include the first phase current, the second phase current, and the third phase current; The sampling time when the judgment value first exceeds the start threshold is marked as the start time; Starting from the start time, the sampling time when the judgment value is continuously lower than the stop threshold for a preset duration is marked as the end time; Several sampling times are reserved before the start time and several sampling times are reserved after the end time to capture the single running segment.

3. The sluice gate operation status monitoring method according to claim 1 or 2, characterized in that, The method further includes: Based on the average value of the three-phase current at all sampling times during the single operation segment, zero-bias correction is performed on the single operation segment; and / or The single running segment is linearly resampled along the time axis, and the linearly resampled single running segment includes a preset number of sampling times.

4. The method for monitoring the operating status of a sluice gate as described in claim 1, characterized in that, When the impact-sensitive characteristic sequence includes an instantaneous imbalance sequence, the step of decoupling the original three-phase current sequence to obtain the impact-sensitive characteristic sequence includes: Based on the instantaneous current value sequence of each phase in the original three-phase current sequence, the sliding effective value is calculated according to the sliding window to obtain the amplitude sequence of each phase; For each sampling time, the average value of the three-phase amplitude is calculated based on the amplitude sequence of each phase as the average value of the three-phase common load water. The standard deviation of the three-phase amplitude relative to the average value of the three-phase common load level is calculated as the three-phase common load level dispersion. Divide the three-phase common load level dispersion by the three-phase common load level average to obtain the instantaneous unbalance value corresponding to the sampling time; The sampling times of the original three-phase current sequence are traversed to obtain the instantaneous unbalance sequence composed of the instantaneous unbalance values.

5. The method for monitoring the operating status of a sluice gate as described in claim 1, characterized in that, When the impact-sensitive characteristic sequence includes a total impact intensity sequence, the step of decoupling the original three-phase current sequence to obtain the impact-sensitive characteristic sequence includes: Based on the instantaneous current value sequence of each phase in the original three-phase current sequence, the instantaneous change intensity value of each sampling time compared with the adjacent sampling time is calculated to obtain the instantaneous change intensity sequence of each phase. The local baseline is calculated based on the instantaneous change intensity sequence of each phase, and the effective impact change value of each phase is obtained by subtracting the local baseline from the instantaneous change intensity value of each phase. The effective impact change value of each phase is fused into three phases to obtain the total impact intensity sequence.

6. The method for monitoring the operating status of a sluice gate as described in claim 1, characterized in that, The method further includes: pre-training the state recognition model; the pre-training of the state recognition model includes: For a single running sample in the running sample set, the impact-sensitive feature sequence of the running sample is extracted to construct a feature matrix, and an initial input feature map is constructed based on the feature matrix; The input feature map is encoded using a convolutional network to obtain an encoded output feature map; the convolutional network includes multiple convolutional blocks of different scales, and a gating network is used to adaptively generate the scale weights of each convolutional block for feature fusion. The predicted probability of each state category is calculated based on the encoded output feature map, and the total loss value is calculated based on the predicted probability. The model parameters of the convolutional network are adjusted using the total loss value until training is complete and the state recognition model is obtained.

7. The sluice gate operation status monitoring method according to claim 6, characterized in that, The total loss value includes a category-weighted cross-entropy loss term and a prototype contrast constraint loss term. The calculation of the predicted probability for each state category based on the encoded output feature map, and the calculation of the total loss value based on the predicted probability, includes: Global average pooling is performed on the encoded output feature map along the pseudo-temporal direction to obtain the sample embedding vector; Calculate the similarity between the sample embedding vector and the category prototype vector corresponding to each state category to obtain the predicted probability of each state category; The category-weighted cross-entropy loss term is calculated based on the predicted probability of each state category; and the prototype comparison constraint loss term is calculated based on the predicted state category corresponding to the maximum predicted probability and the actual state category of the running sample.

8. A sluice gate operation status monitoring device, characterized in that, include: The detection module is used to extract a single operation segment from continuous monitoring data when the opening and closing action of the sluice gate is detected, and to construct the original three-phase current sequence based on the single operation segment; A decoupling module is used to decouple the original three-phase current sequence to obtain an impact-sensitive characteristic sequence, and to construct a feature matrix based on the impact-sensitive characteristic sequence; the impact-sensitive characteristic sequence includes one or more of the following: instantaneous unbalance degree sequence, equivalent positive sequence current amplitude sequence, positive sequence current low-frequency trend sequence, and total impact intensity sequence; The prediction module is used to input the feature matrix into the trained state recognition model, so as to use the state recognition model to perform multi-scale pseudo-temporal coding and category prototype constraint classification, and obtain the predicted probability of each state category output by the state recognition model. The output module is used to take the state category corresponding to the maximum predicted probability as the current sluice gate operating state.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the sluice gate operation status monitoring method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the sluice gate operation status monitoring method as described in any one of claims 1 to 7.

Images