Power management control method and device for power supply of transmission line vibration energy

Abstract

A power management control method and device for power supply of transmission line vibration energy conversion, which converts vibration waveform into a token sequence, inputs a multi-head attention network, only queries the end token to calculate the attention score and take the average value of the whole token key projection to obtain the attention token importance, and screens the sparse active token set to participate in attention calculation; the feedforward network is approximated to low rank, the token low rank transformation amplitude is predicted to obtain the feedforward network token importance, and the sparse active token set of the feedforward network layer is screened according to the transformation calculation; the deep network is divided into subsets, the unified sparse active token set is used in each layer in the subset, the importance score is recalculated based on the final hidden state at the tail layer of the subset to perform pruning, and the sequence length is gradually compressed; finally, the duty cycle of the end token hidden state output is taken to generate a PWM waveform to drive a switching tube, and maximum power point tracking is realized. The present application improves the accuracy of vibration energy conversion MPPT, reduces the reasoning calculation amount, and is suitable for low-power AI chip deployment.

Description

Technical Field

[0001] This invention belongs to the field of power management and control technology, and particularly relates to a power management and control method and device for harvesting energy from vibrations in transmission lines. Background Technology

[0002] The operating environment of power transmission lines is complex, the geographical span of the monitored objects is large, and it is difficult to replace the batteries of the sensors on power transmission lines. While inductive power harvesting can be used on AC power transmission lines, it is difficult to achieve inductive power harvesting on DC power transmission lines, making the power supply problem of the sensors prominent.

[0003] Vibration energy harvesting from transmission lines is a novel micro-energy harvesting technology adapted to the digital transformation of power grids. It primarily targets the low-frequency, micro-amplitude, continuous vibrations of overhead transmission lines caused by wind disturbances and airflow, converting mechanical vibration energy into electrical energy. Compared to traditional power supply methods, vibration energy harvesting from transmission lines offers advantages such as self-powering, strong environmental adaptability, and maintenance-free operation. It can operate in all weather conditions, including rain, snow, cloudy days, and no sunlight, unaffected by day / night cycles or weather changes, effectively solving industry pain points such as the cumbersome replacement of traditional battery power supplies and the weather- and sunlight-limited limitations of photovoltaic energy harvesting.

[0004] Vibration is a common form of energy generated by the natural environment and is applicable to high-voltage transmission lines. Extensive research has been conducted in China, primarily based on three principles: piezoelectric, electromagnetic, and triboelectric nano-power generation. Piezoelectric methods are simple in structure and suitable for medium- to high-frequency vibration scenarios; electromagnetic methods offer stable output power and are suitable for typical line vibration conditions; while triboelectric nano-power generation excels at capturing low-frequency, weak vibrations. For example, the invention patent CN113193786A from the China Electric Power Research Institute proposes a vibration energy harvesting device, a power supply, and a temperature sensor. The vibration energy harvesting device is fixed to the transmission line via a snap-fit ​​connector. As the cantilever beam vibrates with the transmission line, the piezoelectric crystal deforms while following the cantilever beam's vibration, converting the vibration into electrical energy.

[0005] Improving the efficiency of vibration energy harvesting is a key technical issue for self-powered sensors. For example, the invention patent CN120217711A from the Southern Power Grid Research Institute proposes a design method, device, computer equipment, and storage medium for a vibration energy harvester. Based on initial design parameters, it determines the influence ratio of coils, magnets, and springs on key performance parameters, and determines component parameters based on the optimized design sequence, so that the power density of the vibration energy harvester reaches its maximum power density and its characteristic frequency reaches the target characteristic frequency. The invention patent CN113193786A from the China Electric Power Research Institute proposes a multi-directional vibration energy harvesting device to solve the problem of collecting vibration energy from only one direction. Through a frame structure design, it causes the piezoelectric element to deform during vibration to generate electrical energy.

[0006] The aforementioned vibration energy harvesting device has advantages such as simple structure and flexible shape. However, due to the characteristics of vibration energy, the energy directly output by the device is not stable DC power and cannot directly charge the battery or power the sensor. Therefore, an energy harvesting circuit must be designed in the self-harvesting sensor to realize AC-DC conversion and improve the energy harvesting stability in complex environments. Currently, the most typical energy harvesting circuit control algorithm is the maximum power point tracking (MPPT) algorithm. MPPT was first used in photovoltaic cells. Because the intensity of sunlight varies at different times, MPPT is used to track the optimal load in real time, and AI analysis is introduced into the algorithm to maximize the efficiency of photovoltaics. For example, the invention patent CN121433435B of Southwest University proposes a variable coefficient step size maximum power point tracking method for photovoltaic power generation systems. Based on the incremental conductance method, the position of the current operating point relative to the maximum power point is determined, and the adaptive particle swarm optimization algorithm is used to dynamically optimize the adaptive step size factor, which can quickly lock the maximum power point. For example, the invention patent CN119276115B from Harbin Institute of Technology proposes a novel DC energy harvesting and management chip and its operation method. It achieves maximum power point tracking through a transient enhanced MPPT algorithm circuit, enabling high-efficiency management of DC environmental energy such as photovoltaic and thermoelectric power.

[0007] In summary, since the optimal load for vibration energy harvesting varies with the vibration frequency of the input mechanical energy, the maximum electrical energy output can be obtained from the vibration energy harvesting component by tracking and locking the optimal load using MPPT (Maximum Power Point Tracking). However, current transmission line vibration energy harvesting devices generally employ MPPT based on fractional open-circuit voltage (FOCV), optimizing the specific proportional relationship between the voltage under load and the open-circuit voltage by adjusting the period or duty cycle of the switching converter. This method is simple to implement and does not require output feedback, but its drawback is that it is based on empirical values ​​and does not consider the relationship between circuit losses and vibration energy, resulting in an inaccurate maximum power point. There is a need to research methods for efficiently tracking the maximum power point of vibration energy harvesting, improving the efficiency of vibration energy harvesting in transmission lines, and realizing self-powered, environmentally adaptable, and maintenance-free transmission line sensors. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a power management control method and device for power harvesting from transmission line vibrations. The purpose of this invention is to improve the efficiency of transmission line vibration power harvesting management circuits. Addressing the issue of vibration energy variations in the transmission line environment, it utilizes the MPPT method to track the optimal load. To address the need for manual experience in optimizing the cycle or duty cycle of the switching converter, this invention employs a long-sequence inference acceleration method with importance self-predictive layer skipping and post-pruning. Through rapid processing of a large model of the vibration waveform, it achieves accurate locking of the maximum power point, enabling the vibration power harvesting device to maintain stable power output under the complex and variable operating conditions of transmission lines.

[0009] The present invention adopts the following technical solution.

[0010] This invention proposes a power management and control method for energy harvesting from vibration in transmission lines, comprising: One-dimensional long temporal vibration waveforms are converted into word sequences using a temporal convolutional encoder; The word sequence is input into a multi-head attention network, and the attention score between the key projection of all words and the query projection of the last word is calculated. The attention score is then normalized. The arithmetic mean of the last row probabilities obtained by each attention head is taken to obtain the attention word importance. The word with the highest attention word importance is selected to form a sparse activation word set. Only the sparse activation word set participates in the query projection, value projection, and attention calculation. The hidden states of the lexical units in the sparse activated lexical unit set are used as input to the feedforward network layer. A lightweight surrogate subnetwork is introduced to perform a low-rank approximation on the feedforward network. The low-rank transformation magnitude of each lexical unit is predicted by the surrogate subnetwork to obtain the importance of the feedforward network lexical units. Based on the importance of the feedforward network lexical units, a sparse activated lexical unit set for the feedforward network layer is constructed. The feedforward network layer transformation calculation is performed only on the lexical units in the sparse activated lexical unit set of the feedforward network layer. Non-activated lexical units are directly passed through the residual connection. The deep network is divided into several subsets, each subset including several consecutive network layers. All network layers in a subset use the same sparse activation word set. The feedforward network layers of each network layer in a subset use their own generated feedforward network layer sparse activation word sets. After the feedforward network layer of the last network layer in a subset has been calculated, the final importance score is recalculated based on the word hidden state of the final output of the current subset. Word pruning is then performed to form the activation word set of the next subset. After the last layer of the last subset has been computed, the final hidden state of the last word is taken, and the duty cycle is output through the linear regression head to generate a PWM waveform to drive the switching transistor and achieve maximum power point tracking.

[0011] More preferably, the last word refers to the nth word in the word sequence of length n output by the temporal convolutional encoder, corresponding to the latest sliding window of the vibration waveform in the current decoding step.

[0012] More preferably, the importance of the attention lexicon is calculated by taking the arithmetic mean of the last row probabilities obtained from each attention head, and the last row probability is obtained by taking the last row after Softmax normalization of the attention scores between the key projections of all lexicons and the query projections of the last lexicon.

[0013] More preferably, the query projection is obtained by linearly projecting the hidden state of the nth word into the query weight of the hth head, and the key projection is obtained by projecting all words in the sequence into the key weight of the hth head.

[0014] More preferably, the lightweight proxy sub-network is constructed as follows: Singular value decomposition is performed on the two weight matrices of the original feedforward network. The first r singular vectors are used to form low-rank matrices A and B. The surrogate subnetwork is defined as the output after the hidden state of the input token is sequentially transformed by the linear transformation of the low-rank matrix A, the activation function of the original feedforward network, and the linear transformation of the low-rank matrix B. The parameters of the surrogate subnetwork are obtained offline by truncating the singular value decomposition and are not updated with online inference.

[0015] More preferably, the low-rank transformation amplitude is obtained by calculating the L2 norm of the proxy sub-network output, and the importance of the feedforward network lexicon is obtained by multiplying the low-rank transformation amplitude by the importance of the attention lexicon.

[0016] More preferably, the feedforward network output of the last network layer of the subset is normalized to obtain the final hidden state of the subset. During pruning, the final importance score is calculated. The final importance score is obtained by multiplying the L2 norm of the final hidden state of the subset by the attention lexical importance recalculated in the tail layer of the subset, and taking the first... The input lexicon set of the next subset is composed of lexicons.

[0017] More preferably, the pruned words are removed from the sequence, the subsequent positional encodings are rearranged according to the new sequence, and the first layer of the multi-head attention network of the next subset only takes the compressed sequence as input.

[0018] More preferably, the duty cycle Here, S1 represents the duty cycle of the switch in the buck-boost circuit, and S2 has a duty cycle of 1. , This is a single value output by the linear regression head.

[0019] This invention also proposes a control device for power transmission line vibration energy harvesting, used for maximum power point tracking control of power transmission line vibration energy harvesting components. The power transmission line vibration energy harvesting components include a vibrating mass block, a piezoelectric element, and structural components. The power transmission line vibration energy harvesting control device includes a vibration energy harvesting component, an energy harvesting circuit, a control circuit, a wireless communication component, a sensor signal sampling circuit, a host terminal, and an AI acceleration chip. The energy harvesting circuit includes a rectifier, a switching converter, and an energy storage element; The host computer stores a computer program. The AI ​​acceleration chip implements the steps of the above method when executing the computer program.

[0020] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps according to the method described above.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention addresses the problem of vibration energy variation in the transmission line environment by using a maximum power point tracking method to track and lock the optimal load, enabling the vibration energy harvesting device to maintain stable power output under the complex and variable operating conditions of the transmission line.

[0022] 2. This invention addresses the problem of requiring manual adjustment of the cycle or duty cycle of the switching converter. It uses a large model to analyze the vibration model and employs a long-sequence inference acceleration method based on importance self-predictive skipping and post-pruning to improve speed and reduce the power consumption of the control algorithm. This is beneficial for realizing an intelligent self-powered transmission line monitoring device.

[0023] 3. Comparative experiments were conducted on a standard vibration test bench. Under typical light wind vibration conditions of 25Hz and 2mm amplitude, the MPPT locking accuracy of this invention reached 96.8%, the single inference time was 3.2ms, the peak computing power consumption of the AI ​​chip was 2.3 TOPS, and the sequence compression ratio reached 5.8 times. In the full-frequency scan experiment (10Hz~150Hz), this invention maintained an MPPT locking accuracy of over 92% at all test frequencies, with a more significant advantage in the high-frequency range (above 100Hz). The above experimental results show that this invention can significantly reduce the inference computation and computing power consumption while ensuring MPPT locking accuracy, making it suitable for low-power AI chip deployment. Attached Figure Description

[0024] Figure 1 This is a flowchart of a power management control method for harvesting energy from vibration in transmission lines according to the present invention. Figure 2 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 3 This is a flowchart of step 101 in Embodiment 1 of the present invention; Figure 4 This is a flowchart of step 103 in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the unidirectional vibration energy harvesting component for transmission lines according to the present invention; Figure 6 This is a schematic diagram of the multi-directional vibration energy harvesting component for power transmission lines according to the present invention; Figure 7 This is a connection diagram of the power transmission line vibration energy harvesting control device of the present invention; Figure 8 This is a schematic diagram of the energy harvesting circuit in the power transmission line vibration energy harvesting control device of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0026] like Figure 1 As shown, this invention proposes a power management and control method for energy harvesting from transmission line vibration, specifically including the following steps: One-dimensional long temporal vibration waveforms are converted into word sequences using a temporal convolutional encoder; The word sequence is input into a multi-head attention network, and the attention score between the key projection of all words and the query projection of the last word is calculated. The attention score is then normalized. The arithmetic mean of the last row probabilities obtained by each attention head is taken to obtain the attention word importance. The word with the highest attention word importance is selected to form a sparse activation word set. Only the sparse activation word set participates in the query projection, value projection, and attention calculation. The term "last word" refers to the nth word in the word sequence of length n output by the temporal convolutional encoder, corresponding to the latest sliding window of the vibration waveform in the current decoding step.

[0027] The importance of the attention lexicon is calculated by taking the arithmetic mean of the last row probabilities obtained from each attention head. The last row probability is obtained by taking the last row after Softmax normalization of the attention scores between the key projections of all lexicons and the query projections of the last lexicon.

[0028] The query projection is obtained by linearly projecting the hidden state of the nth word through the query weight of the hth head, and the key projection is obtained by projecting all words in the sequence through the key weight of the hth head.

[0029] The hidden states of the lexical units in the sparse activated lexical unit set are used as input to the feedforward network layer. A lightweight surrogate subnetwork is introduced to perform a low-rank approximation on the feedforward network. The low-rank transformation magnitude of each lexical unit is predicted by the surrogate subnetwork to obtain the importance of the feedforward network lexical units. Based on the importance of the feedforward network lexical units, a sparse activated lexical unit set for the feedforward network layer is constructed. The feedforward network layer transformation calculation is performed only on the lexical units in the sparse activated lexical unit set of the feedforward network layer. Non-activated lexical units are directly passed through the residual connection. The lightweight proxy sub-network is constructed as follows: Singular value decomposition is performed on the two weight matrices of the original feedforward network. The first r singular vectors are used to form low-rank matrices A and B. The surrogate subnetwork is defined as the output after the hidden state of the input token is sequentially transformed by the linear transformation of the low-rank matrix A, the activation function of the original feedforward network, and the linear transformation of the low-rank matrix B. The parameters of the surrogate subnetwork are obtained offline by truncating the singular value decomposition and are not updated with online inference.

[0030] The low-rank transformation amplitude is obtained by calculating the L2 norm of the proxy sub-network output, and the importance of the feedforward network lexicon is obtained by multiplying the low-rank transformation amplitude by the importance of the attention lexicon.

[0031] The deep network is divided into several subsets, each subset including several consecutive network layers. All network layers in a subset use the same sparse activation word set. The feedforward network layers of each network layer in a subset use their own generated feedforward network layer sparse activation word sets. After the feedforward network layer of the last network layer in a subset has been calculated, the final importance score is recalculated based on the word hidden state of the final output of the current subset. Word pruning is then performed to form the activation word set of the next subset. The feedforward network output of the last network layer of the subset is normalized to obtain the final hidden state of the subset. During pruning, the final importance score is calculated. The final importance score is obtained by multiplying the L2 norm of the final hidden state of the subset by the attention lexical importance recalculated in the last layer of the subset, and taking the first... The input lexicon set of the next subset is composed of lexicons.

[0032] The pruned words are removed from the sequence, and the subsequent positional encodings are rearranged according to the new sequence. The first layer of the multi-head attention network in the next subset only takes the compressed sequence as input.

[0033] After the last layer of the last subset has been computed, the final hidden state of the last word is taken, and the duty cycle is output through the linear regression head to generate a PWM waveform to drive the switching transistor and achieve maximum power point tracking.

[0034] The duty cycle Here, S1 represents the duty cycle of the switch in the buck-boost circuit, and S2 has a duty cycle of 1. , This is a single value output by the linear regression head.

[0035] This invention also proposes a control device for power transmission line vibration energy harvesting, used for maximum power point tracking control of power transmission line vibration energy harvesting components. The power transmission line vibration energy harvesting components include a vibrating mass block, a piezoelectric element, and structural components. The power transmission line vibration energy harvesting control device includes a vibration energy harvesting component, an energy harvesting circuit, a control circuit, a wireless communication component, a sensor signal sampling circuit, a host terminal, and an AI acceleration chip. The energy harvesting circuit includes a rectifier, a switching converter, and an energy storage element; The host computer stores a computer program. The AI ​​acceleration chip implements the steps of the above method when executing the computer program.

[0036] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps according to the method described above.

[0037] Example 1 This invention proposes a power management and control method for energy harvesting from transmission line vibrations. Addressing the complex and variable vibration signals in transmission line environments, it proposes a long-sequence inference acceleration method based on importance self-predictive layer skipping and post-pruning. By analyzing the characteristics of multi-head attention and feedforward networks, and through rapid processing of a large model of the vibration waveform, it achieves accurate locking of the maximum power point, including: 1) A one-dimensional long temporal vibration waveform is converted into a word sequence by a temporal convolutional encoder, and a sparse activation word set is obtained based on local attention to accelerate inference; 2) Construct a sparse set of activated words for the feedforward network layer, skip the calculation of non-activated words during the inference process, and improve the inference efficiency of the feedforward network layer; 3) Divide the deep network into several subsets, prune based on the importance of attention lexicons, gradually compress the sequence length, and quickly lock the maximum power point.

[0038] The basic process is as follows Figure 2As shown, the control method for energy harvesting from transmission line vibration includes the following steps: Step 101: Accelerating Vibration Waveform Lexical Sequence Inference Based on Local Attention A sensor signal sampling circuit obtains a one-dimensional long-sequence vibration waveform, which is then converted into a word sequence of length n by a temporal convolutional encoder. This word sequence is input into a multi-head attention network (MHA). Within this network, the attention score is normalized using a softmax function. The normalized attention between the key projections of all words and the query projections of the last word is calculated, and the average value is taken to obtain the local attention importance, guiding the hop selection of the MHA network.

[0039] in, For attention lemma importance, A represents the total number of attention heads in the multi-head attention network, and D represents the feature dimension. This represents the key projection of all lexical units in the h-th attention head. This represents the query projection of the last word in the h-th attention head. Then, the most important words are selected to form a sparse activation word set:

[0040] Only the sparse activation lexicon set (SA) participates in query, value projection, and attention calculations. This significantly reduces the computational complexity of the current layer while compressing the lexicon cache, further reducing the inference overhead in the decoding stage. The process is as follows: Figure 3 As shown.

[0041] In this embodiment, "last word" refers to the nth word in a word sequence of length n output by the temporal convolutional encoder, corresponding to the latest sliding window of the vibration waveform in the current decoding step; query projection The weight is determined by querying the hidden state of the nth term through the hth head. Linear projection yields key projection. From the sequence of all morphemes Obtained by projection. Take before SA is composed of 1 word element. Take a value of 0.3 to 0.5, and decrease by 0.1 for each subset as the depth of the subset increases.

[0042] While existing technologies (such as SpAtten) propose using attention probability as lexical importance, they involve summing all query positions, accumulating across headers, and recalculating layer by layer, resulting in high computational costs and unsuitability for the "decision of the next step based on the current step state" requirement in vibration waveform streaming decoding. This step 101 transforms the importance metric from "global historical attention" to "current step decision attention" by using only the last lexical query and averaging across headers, making it more suitable for MPPT scenarios and laying the foundation for subsequent acceleration.

[0043] Step 102: Accelerating with a feedforward network based on lexical importance The hidden states of the lexical units in the sparse activation lexical set SA output in step 101 are used as input to the feedforward network (FFN) layer.

[0044] To achieve efficient online lexical importance prediction, a lightweight proxy subnetwork is introduced to perform a low-rank approximation on the feedforward network.

[0045] The lightweight proxy subnetwork is constructed as follows: Singular value decomposition (SVD) is performed on the two-layer weight matrices of the original FFN, and the first r singular vectors are used to form low-rank matrices A and B. The proxy network is defined as follows: ,in , This is the activation function of the original FFN. The parameters of this agent subnetwork are obtained offline through SVD truncation and are not updated with online inference.

[0046] The original FFN was , .right Perform SVD and take the first r left singular vectors to construct ,right Perform SVD and take the first r left singular vectors to construct Proxy network ,in (Typical value range is) to This strikes a balance between approximate accuracy and computational savings. This value is determined by testing the FFN output approximation error at different r values ​​on a validation set to ensure negligible impact on MPPT locking accuracy. Low-rank transform amplitude Used to estimate the state update amount of this term in the FFN. FFN layer importance is taken as... Top-k (k is the same as MHA layer) )become ;No Lexical skipping The calculation involves directly inputting the MHA output plus the residual into LayerNorm, i.e. This applies to this type of word.

[0047] Predict the low-rank transformation magnitude of each word using a proxy subnetwork. Constructing the importance of lexical units in a feedforward network (or merged with the attention lexical importance in step 101) Based on the importance of the lexical units in the feedforward network, the preceding lexical units are selected. A set of sparse activation lexes is constructed for the feedforward network layer. ,in The value is set to 0.3 to 0.5, decreasing by 0.1 with each subset depth, where n is the number of lexical units in the current subset. A set of lexical units with high importance in the feedforward network is selected to construct a sparse activation lexical set for the feedforward network layer. In the reasoning process, only for The tokens in the feedforward network perform complete FFN two-layer transformation calculations, and the inactive tokens are directly passed through the residual connections, thereby skipping the calculation of inactive tokens and improving the inference efficiency of the feedforward network layer.

[0048] In step 102, the importance of the feedforward network terminology is obtained by multiplying the low-rank transform amplitude predicted by the surrogate subnetwork by the importance of the attention terminology from step 101. The low-rank transform amplitude reflects the amount of information a terminology provides in the nonlinear transform of the feedforward network, while the importance of the attention terminology reflects the relational value of the terminology in the sequence to the current decoding step. Multiplying the two is equivalent to jointly evaluating the comprehensive contribution of the terminology in both the MHA and FFN modules. The value of the SVD truncation parameter r of the surrogate subnetwork strikes a balance between approximate accuracy and computational savings. The scheduling strategy of decreasing α with subset depth ensures that the early subset retains more terms to maintain contextual information, while the later subset is more aggressively compressed to reduce computational load, which matches the physical characteristics of the vibration waveform MPPT: "sufficient perception of vibration state in the early stage and rapid decision-making in the later stage."

[0049] Step 103: Lock the maximum power point like Figure 4 As shown, deep networks are divided into several subsets:

[0050] in, This represents the 1 to t consecutive network layers that constitute the x-th subset, where all network layers within the subset use the unified sparse activation term set SA generated in step 101; This represents the t-th network layer in the x-th subset. Meanwhile, the FFN layers of each layer within the subset use the layers generated in step 102 respectively. Perform FFN calculations to skip levels.

[0051] subset Tail layer The FFN output, after being processed by LayerNorm, yields the final hidden state of the subset. During pruning, the final importance score is calculated by comprehensively considering the state update magnitude of the lexical units and the attention importance.

[0052] in, The L2 norm of the hidden state reflects the amount of information of the lexical after processing by the complete subset; At the tail level of the subset In the middle, the attention importance of the last query across the head is recalculated according to the method in step 101. Take the first... The input lexicon set that constitutes the next subset consists of lexicons. :

[0053] in It ranges from 0.3 to 0.5, decreasing by 0.1 with each subset depth.

[0054] The pruned terms are removed from the sequence, and subsequent positional encodings are rearranged according to the new sequence (this does not affect relative positional encoding or TCN fixed convolutional kernels). The first layer of the next subset of MHA uses only this compressed sequence as input.

[0055] Through the above steps, this invention constructs a "coarse-to-fine" sequence compression pipeline: Step 101 quickly filters out most unimportant terms at the MHA layer; Step 102 further eliminates terms that contribute little to the nonlinear transformation at the FFN layer; and Step 103 performs a thorough sequence compression at the end of the subset composed of multiple layers. These three steps work together to reduce the inference computation of the Transformer to one-tenth or even less of its original amount while maintaining the MPPT locking accuracy, making it possible to run large models on low-power AI chips.

[0056] In step 103, a unified sparse activation lexicon set SA is used in each layer within the subset, avoiding the additional overhead of recalculating attention importance layer by layer. The pruning of the last layer of the subset is based on the product of the L2 norm of the final hidden state and the importance of the attention lexicon. This fusion score takes into account both the information content of the lexicon after processing the complete subset and its attention contribution to the current decoding step. The scheduling strategy, which decreases with the depth of the subset, is consistent with the value selection logic in steps 101 and 102, ensuring a smooth transition in sequence compression. After the pruned tokens are removed from the sequence, the subsequent positional encodings are rearranged according to the new sequence. The first layer MHA of the next subset only takes the compressed sequence as input, thereby achieving progressive compression of the sequence length.

[0057] Step 104: Lock the maximum power point and generate a control signal After the final layer of the last subset has been computed, the final hidden state of the last word is taken, and a single value is output through the linear regression head. The duty cycle of switch S1 in the buck-boost circuit is used as the duty cycle; the control circuit generates a PWM waveform based on the duty cycle to drive S1 and S2, where the duty cycle of S2 is... This allows for adjustment of the output voltage, thereby achieving maximum power point tracking.

[0058] Example 2 The technical solution of the present invention is applicable to a non-intrusive self-powered voltage sensor for 500kV high-voltage transmission lines. The non-intrusive self-powered voltage sensor includes a sensing component, a vibration-powered component, and a power harvesting control device.

[0059] The power transmission line vibration energy harvesting component includes: a vibrating mass block, a piezoelectric element, and a structural component. The structural component is fixedly connected to the power transmission line. One end of a cantilever beam is fixedly connected to the structural component. Both the upper and lower surfaces of the cantilever beam have a layer of piezoelectric crystals. A mass block is fixed to one end of the cantilever beam. As the cantilever beam vibrates with the power transmission line, the piezoelectric crystals deform during the vibration, converting pressure into electrical energy. Upper and lower limit springs are installed on the structural component above and below the mass block, respectively, to reduce energy loss from the cantilever beam and prevent breakage. Figure 5 As shown.

[0060] In another embodiment, the multi-directional vibration energy harvesting component of the transmission line includes: a vibrating mass block, piezoelectric elements, and structural components. The structural components enclose a receiving space and are snap-fitted and fixed to the transmission line. The vibrating mass block is disposed within the receiving space and fixedly connected to the structural components. Piezoelectric elements are provided on the surfaces of the connecting parts. When the structural components are subjected to vibration, the vibration is transmitted to the vibrating mass block, causing the piezoelectric elements to deform and generate electrical energy. Figure 6 As shown.

[0061] The transmission line vibration energy harvesting control device includes a vibration energy harvesting component, an energy harvesting circuit, a control circuit, a wireless communication component, a sensor signal sampling circuit, a main unit, and an AI acceleration chip. The vibration energy harvesting component is connected to the energy harvesting circuit. The control circuit, energy harvesting circuit, sensor signal sampling circuit, and wireless communication component are connected. The wireless communication components are wirelessly connected to each other. The wireless communication component is connected to the main unit and the AI ​​acceleration chip. The long-sequence vibration waveform is obtained through the sensor signal sampling circuit and transmitted to the main unit and AI acceleration chip located on the ground through the wireless communication component. Figure 7 As shown.

[0062] The energy harvesting circuit includes a rectifier, a switching converter, and an energy storage element. The rectifier, composed of diodes, converts the input AC voltage of the vibration energy harvesting component into the output DC voltage. CS is a supercapacitor for energy storage. The buck-boost switching converter consists of two MOSFETs, S1 and S2, and an inductor LS. RL is the equivalent load of the sensor signal sampling circuit, control circuit, and wireless communication components. CF is a filter capacitor. The voltage of RL is kept stable by adjusting the duty cycles of S1 and S2. Figure 8 As shown.

[0063] The host device stores a computer program. When the AI ​​acceleration chip executes the computer program, it implements the specific steps of the control algorithm described in Embodiment 1. After locking the maximum power point, it sends the data to the control circuit via a wireless communication component to adjust the period or duty cycle of the switching converter.

[0064] Specifically, the model in the AI ​​acceleration chip is a Transformer trained offline on a dataset collected on a vibration table. The dataset covers the typical vibration frequency range (5Hz~200Hz) and different amplitude levels of transmission lines. Each sample includes the vibration waveform, open-circuit voltage, and the optimal duty cycle calibrated using a frequency sweep method, totaling no fewer than 100,000 samples. The model outputs a single value in the AI ​​acceleration chip. This value is directly used as the duty cycle of switch S1 in the buck-boost circuit. After receiving this duty cycle signal, the control circuit generates a corresponding PWM waveform to drive S1 and S2 (S2 duty cycle). This allows for adjustment of the output voltage, thereby achieving maximum power point tracking.

[0065] Example 3 The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps according to the method described above.

[0066] Example 4 To verify the technical effectiveness of this invention, a comparative experiment was conducted on a standard vibration test bench. The experimental conditions were as follows: Experimental platform: Vibration table: frequency range 5Hz~200Hz, amplitude range 0.1mm~5mm; Energy extraction circuit: buck-boost converter, input voltage range 5V~30V, rated power 2W; Model: Transformer-4 layers, hidden dimension 256, FFN intermediate dimension 1024, attention head 8.

[0067] The control group setup is shown in Table 1: Table 1 Control group setup

[0068] The experimental results (vibration frequency 25Hz, amplitude 2mm, typical light wind vibration condition) are shown in Table 2: Table 2 shows the experimental results.

[0069] According to Table 2, the MPPT locking accuracy of this invention reaches 96.8%, significantly better than SpAtten+SPTS's 93.2%. The analysis is as follows: The last query cross-head average focuses more on the power status judgment of the "current decoding step" rather than SpAtten's "historical cumulative attention"; Subset post-pruning avoids the "importance oscillation" caused by layer-by-layer recalculation, making sequence compression smoother.

[0070] This invention reduces inference time by approximately 43% and computational power consumption by 44% compared to SpAtten+SPTS. The reasons are analyzed as follows: Unifying the SA set within a subset reduces the frequency of calls to the intermediate Top-k engine; Non-activated terms directly pass through residuals, avoiding unnecessary calculations in the FFN layer.

[0071] The sequence compression ratio of this invention reaches 5.8 times, which is higher than the 4.5 times of SpAtten+SPTS. The subset tail layer fusion scoring (hidden state L2 norm × attention importance) of this invention is more robust than single attention scoring and can perform more aggressive pruning without sacrificing accuracy.

[0072] The results of the vibration frequency scanning experiment (amplitude 2 mm, frequency 10 Hz ~ 150 Hz) are shown in Table 3: Table 3. Results of the vibration frequency sweep experiment

[0073] As shown in Table 3, the present invention maintains an MPPT locking accuracy of over 92% at all test frequencies, and its advantage is more obvious in the high-frequency range (above 100Hz) (more than 5 percentage points higher than SpAtten+SPTS), indicating that the "last query + subset post-pruning" strategy has better adaptability to rapid changes in vibration waveforms.

[0074] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0075] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0076] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0077] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A power management and control method for energy harvesting from vibration in transmission lines, characterized in that, include: One-dimensional long temporal vibration waveforms are converted into word sequences using a temporal convolutional encoder; The word sequence is input into a multi-head attention network, and the attention score between the key projection of all words and the query projection of the last word is calculated. The attention score is then normalized. The importance of attentional terms is obtained by taking the arithmetic mean of the probabilities of the last row obtained from each attention head; The sparse activation lexicon set is formed by selecting the lexicon with the highest attention lexicon importance, and only the sparse activation lexicon set participates in query projection, value projection and attention calculation; The hidden states of the lexical units in the sparse activated lexical unit set are used as input to the feedforward network layer. A lightweight surrogate subnetwork is introduced to perform a low-rank approximation on the feedforward network. The low-rank transformation magnitude of each lexical unit is predicted by the surrogate subnetwork to obtain the importance of the feedforward network lexical units. Based on the importance of the feedforward network lexical units, a sparse activated lexical unit set for the feedforward network layer is constructed. The feedforward network layer transformation calculation is performed only on the lexical units in the sparse activated lexical unit set of the feedforward network layer. Non-activated lexical units are directly passed through the residual connection. The deep network is divided into several subsets, each subset including several consecutive network layers. All network layers in a subset use the same sparse activation word set. The feedforward network layers of each network layer in a subset use their own generated feedforward network layer sparse activation word sets. After the feedforward network layer of the last network layer in a subset has been calculated, the final importance score is recalculated based on the word hidden state of the final output of the current subset. Word pruning is then performed to form the activation word set of the next subset. After the last layer of the last subset has been computed, the final hidden state of the last word is taken, and the duty cycle is output through the linear regression head to generate a PWM waveform to drive the switching transistor and achieve maximum power point tracking.

2. The power management and control method for energy harvesting from vibration in transmission lines according to claim 1, characterized in that: The term "last word" refers to the nth word in the word sequence of length n output by the temporal convolutional encoder, corresponding to the latest sliding window of the vibration waveform in the current decoding step.

3. The power management and control method for energy harvesting from vibration in transmission lines according to claim 1, characterized in that: The importance of the attention lexicon is calculated by taking the arithmetic mean of the last row probabilities obtained from each attention head. The last row probability is obtained by taking the last row after Softmax normalization of the attention scores between the key projections of all lexicons and the query projections of the last lexicon.

4. The power management and control method for energy harvesting from vibration in transmission lines according to claim 1, characterized in that: The query projection is obtained by linearly projecting the hidden state of the nth word into the hth head query weight, and the key projection is obtained by projecting all words in the sequence into the hth head key weight.

5. The power management and control method for energy harvesting from vibration in transmission lines according to claim 1, characterized in that: The lightweight proxy sub-network is constructed as follows: Singular value decomposition is performed on the two weight matrices of the original feedforward network. The first r singular vectors are used to form low-rank matrices A and B. The surrogate subnetwork is defined as the output after the hidden state of the input token is sequentially transformed by the linear transformation of the low-rank matrix A, the activation function of the original feedforward network, and the linear transformation of the low-rank matrix B. The parameters of the surrogate subnetwork are obtained offline by truncating the singular value decomposition and are not updated with online inference.

6. The power management and control method for energy harvesting from vibration in transmission lines according to claim 5, characterized in that: The low-rank transformation amplitude is obtained by calculating the L2 norm of the proxy sub-network output, and the importance of the feedforward network lexicon is obtained by multiplying the low-rank transformation amplitude by the importance of the attention lexicon.

7. The power management and control method for energy harvesting from vibration in transmission lines according to claim 1, characterized in that: The feedforward network output of the last network layer of the subset is normalized to obtain the final hidden state of the subset. During pruning, the final importance score is calculated. The final importance score is obtained by multiplying the L2 norm of the final hidden state of the subset by the attention lexical importance recalculated in the last layer of the subset, and taking the first... The input lexicon set of the next subset is composed of lexicons.

8. The power management and control method for energy harvesting from vibration in transmission lines according to claim 1, characterized in that: The pruned words are removed from the sequence, and the subsequent positional encodings are rearranged according to the new sequence. The first layer of the multi-head attention network of the next subset only takes the compressed sequence as input.

9. A power management and control method for energy harvesting from vibration in transmission lines according to claim 1, characterized in that: The duty cycle Here, S1 represents the duty cycle of the switch in the buck-boost circuit, and S2 has a duty cycle of 1. , This is a single value output by the linear regression head.

10. A control device for power transmission line vibration energy harvesting using the method of any one of claims 1-9, used for maximum power point tracking control of the power transmission line vibration energy harvesting component, wherein the power transmission line vibration energy harvesting component comprises a vibrating mass block, a piezoelectric element, and a structural component, characterized in that: The power transmission line vibration energy harvesting control device includes a vibration energy harvesting component, an energy harvesting circuit, a control circuit, a wireless communication component, a sensor signal sampling circuit, a host terminal, and an AI acceleration chip. The energy harvesting circuit includes a rectifier, a switching converter, and an energy storage element; The host computer stores a computer program. When the AI ​​acceleration chip executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

11. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1 to 9.

Images