A magnet power supply control system

Abstract

The present invention provides a magnet power supply control system, which relates to the field of data processing technology. The system is used to: obtain parameter setting instructions input by a control terminal, determine the control input parameters at each time point based on the set current value and the actual voltage value, construct a second data matrix based on the historical operation sequence index value, and superimpose the first data matrix and the second data matrix to obtain a fused data tensor. Based on the fused data tensor, gradient features are extracted, and the response fit between the gradient features and each control path is calculated to obtain a pattern fit index set. The maximum response fit is extracted and the state transition direction is determined to obtain a state transition index. The control parameter mapping path is called, and the logical conditions of the physical circuit change are mapped into a digital control sequence to obtain a control factor matrix. The execution instruction is generated and sent to the power output execution layer. The present invention can replace hardware control with data drive to ensure the stable operation of the magnet power supply.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, in particular to a magnet power supply control system. Background Art

[0002] In nuclear medicine radiotherapy equipment, magnet power control systems provide precise excitation current to beam control magnets, thereby controlling the trajectory of charged particles and adjusting the beam's direction and convergence characteristics. Existing systems typically utilize a central control console in conjunction with a front-end controller. Operating parameters are set through an operator interface, and the controller outputs these parameters to the power module, forming a complete power control chain. These systems rely on hardware circuitry to implement the control loop. Adjusting parameters or changing control strategies requires replacing circuit components or adjusting hardware connections, which can lead to system downtime or electrical interface incompatibility.

[0003] For example, the system that controls the scanning magnet in a large radiotherapy device currently uses a periodic sampling mode, which polls the current output status at a fixed period and makes adjustments. However, this mode may have a response lag when dealing with emergencies such as abnormal cooling water flow and sudden changes in coil temperature. After the sensor detects an abnormal temperature or flow signal, the system may delay the execution of the power-off protection operation, causing beam deviation or damage to the power module. Summary of the Invention

[0004] The object of the present invention is to provide a magnet power supply control system, aiming to solve the problems mentioned in the background technology.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0006] A magnet power supply control system, the system comprising:

[0007] An input module is used to obtain parameter setting instructions and operation mode identifiers input by the control terminal, and extract the set current value, actual voltage value and historical operation sequence index value;

[0008] An aggregation module is used to determine the control input parameters at each time point based on the set current value and the actual voltage value, and construct a first data matrix; construct a second data matrix based on the historical operation sequence index value, and superimpose the first data matrix and the second data matrix to obtain a fused data tensor;

[0009] The fitting module is used to extract gradient features based on the fused data tensor, obtain a feature offset vector group, and calculate the response fit between it and each control path to obtain a pattern fitting index set;

[0010] A conversion module is used to extract the maximum response fit and determine the state conversion direction according to the operation mode identifier and the mode fitting index set, thereby obtaining a state conversion index;

[0011] A mapping module is used to call the control parameter mapping path according to the state conversion index, map the logical conditions of the physical circuit change into a digital control sequence, and obtain the control factor matrix;

[0012] The control module is used to generate execution instructions according to the control factor matrix and send them to the power output execution layer.

[0013] Furthermore, the aggregation module includes:

[0014] A sampling unit is used to extract control input parameters according to corresponding combinations of set current values ​​and actual voltage values ​​at each time point to obtain an input feature vector group;

[0015] An input matrix construction unit is used to arrange the input feature vector group into a two-dimensional array in time sequence, and use the input feature vectors as columns and the time points as rows to obtain a first data matrix;

[0016] A sequence mapping unit is used to extract control configuration parameters according to the historical operation sequence index value to obtain a historical configuration vector group;

[0017] The sequence matrix construction unit is used to arrange the historical configuration vector group into a two-dimensional array in index order, and use the historical configuration vectors as columns and the historical operation sequences as rows to obtain a second data matrix.

[0018] Furthermore, the aggregation module further includes:

[0019] A sample matching unit is used to extract each column of input feature vectors from the first data matrix and each column of historical configuration vectors from the second data matrix to obtain an input configuration sample pair set;

[0020] A matching fusion scoring unit is used to match each pair of vectors in the input configuration sample pair set and calculate the matching fusion score of the sample pair;

[0021] The matching matrix unit is used to arrange the matching fusion scores according to the input feature vector number and the historical configuration vector number to obtain a two-dimensional matching matrix set, and use the two-dimensional matching matrix at each time point as an independent matching channel to obtain a three-dimensional fusion structure;

[0022] The fusion tensor construction unit is used to stack all independent matching channels in the three-dimensional fusion structure in chronological order, determine the fusion score of the input feature vector and the historical configuration group at the corresponding time point, and obtain the fusion data tensor.

[0023] Furthermore, the matching fusion scoring unit includes:

[0024] The matching fusion score calculation unit is used to perform feature mapping on the input feature vector to obtain a feature mapping value; and to perform normalization processing on the historical configuration vector to obtain a normalized strength value;

[0025] According to the feature map value and the normalized intensity value, the difference is calculated and nonlinearly amplified to obtain the fused difference term; the fused difference terms of all input feature vectors and historical configuration vectors are double-summed to obtain the total fused difference term;

[0026] The configuration response accumulation item is calculated based on the product of the historical configuration vector and the normalized intensity value; the total fusion difference item and the configuration response accumulation item are fused to obtain the fusion matching factor; the fusion matching factor is normalized to obtain the matching fusion score.

[0027] Furthermore, the fitting module includes:

[0028] A gradient extraction unit is used to calculate the response change rate of the input feature vector under different historical configuration vectors based on the fused data tensor to obtain a gradient tensor set;

[0029] The offset vector unit is used to compress the dimension of the gradient tensor set, aggregate it along the time axis, calculate the response equilibrium offset of each input feature vector under all historical configuration vectors, and obtain the feature offset vector group;

[0030] A control path mapping unit is used to determine the characteristic response structure of the input characteristics corresponding to different control logics according to the control path, and obtain a path feature vector set;

[0031] The fitting index unit is used to map the characteristic offset vector group and the path characteristic vector set, calculate the response fitting degree, and obtain the pattern fitting index set.

[0032] Furthermore, the offset vector unit includes:

[0033] The temporal aggregation unit is used to average the gradient values ​​of the same input feature dimension at all time points in the gradient tensor set to obtain the response mean matrix;

[0034] The configuration alignment unit is used to perform structural matching on each row in the response mean matrix and the historical configuration vector number, and uniformly map them to the corresponding historical configuration dimension to obtain the configuration aggregation matrix;

[0035] An offset calculation unit is used to perform standard deviation compression and normalization on the response mean of each input feature dimension in the configuration aggregation matrix to obtain a response equalization offset value;

[0036] The vector generation unit is used to arrange the response equalization offset values ​​of all input feature dimensions into a vector structure in dimensional order to obtain a feature offset vector group.

[0037] Furthermore, the fitting index unit includes:

[0038] A response fitting calculation unit is used to calculate the characteristic mean of the path characteristic vector under different control paths;

[0039] Based on the offset eigenvector and the path eigenvector of the input eigenvector, the numerical projection ratio of the two on each input feature dimension is calculated to obtain a weighted projection term; based on the absolute difference between the offset eigenvector and the path eigenvector, a symmetric normalized ratio is constructed to obtain a nonlinear suppression term; the weighted projection term and the nonlinear suppression term are multiplied item by item according to the dimension number and summed to obtain the total path fitting term;

[0040] The variance of the path eigenvector is calculated, and the stability penalty term is constructed based on it and the eigenmean. The total path fitting term is used as the numerator and the stability penalty term as the denominator to obtain the response fitting degree.

[0041] Furthermore, the conversion module includes:

[0042] The matching extraction unit is used to traverse and compare all response fitting degrees in the pattern fitting index set, extract the control path number with the largest fitting degree value, and obtain the maximum fitting path index;

[0043] The mode recognition unit is used to filter out the control path index corresponding to the current operation mode according to the operation mode identifier, and compare it with the maximum fitting path index to obtain the target path number;

[0044] The state direction unit is used to extract the corresponding path feature vector according to the target path number, and perform difference calculation with the feature offset vector of the corresponding dimension to obtain a response difference vector. Then, based on the response difference vector, the adjustment direction of the current input feature is determined to obtain the state offset direction;

[0045] The index generating unit is used to encode and map the maximum fitting path index with the state offset direction to obtain the state conversion index.

[0046] Furthermore, the state direction unit includes:

[0047] A feature extraction unit is used to extract the path feature vector with the corresponding number from the path feature vector set according to the target path number to obtain the target path vector;

[0048] a difference construction unit, configured to perform dimension-by-dimension difference operation on the target path vector and the offset vector of the corresponding dimension in the feature offset vector group to obtain a response difference vector;

[0049] A symbol parsing unit is used to parse the symbol information of each dimension in the response difference vector, mark the offset direction of the current input feature compared to the control, and obtain an adjustment direction set;

[0050] The trend aggregation unit is used to count the distribution trend of the overall offset direction according to the adjustment direction set, and determine a unified direction to obtain the state offset direction.

[0051] Furthermore, the mapping module includes:

[0052] The path calling unit is used to match the state conversion index with the control parameter mapping path to obtain the parameter configuration path;

[0053] A conditional parsing unit is used to read the circuit logic change conditions in the parameter configuration path and parse the module change signals therein to obtain a logic tuple sequence;

[0054] a digital generation unit for combining and encoding the module change signal and the control path number in the logic tuple sequence to obtain a digital control sequence;

[0055] The factor matrix unit is used to expand the sequence data into a two-dimensional control matrix structure according to the digital control sequence to obtain a control factor matrix.

[0056] The above solution of the present invention includes at least the following beneficial effects:

[0057] The present invention generates a first data matrix by setting current values ​​and actual voltage values ​​to reflect the dynamic state of real-time control inputs. At the same time, a second data matrix is ​​constructed based on historical operation sequences to represent historical configuration modes or control experience. By superimposing matrices, a fused data tensor is generated to achieve temporal integration and structural consistency mapping of data. This fusion not only retains the sensitivity of real-time parameters but also introduces the stability characteristics of historical operations, thereby improving the system's fault tolerance and predictability for complex state changes. Compared to traditional single-parameter control methods, this aggregated structure has multi-dimensional correlation, high redundancy, and good scalability, which helps to achieve a more robust control decision-making mechanism.

[0058] The present invention extracts feature gradients by fusing data tensors, generates feature offset vectors, and calculates matching fit based on the control path response structure, thereby realizing a data-driven response path selection mechanism. It can automatically extract response difference characteristics of control behaviors under different paths based on the current fused data, construct a set of pattern fitting indicators in high-dimensional space, and realize quantitative expression of the system's adaptability to control logic. It significantly improves the adaptive ability of the control system, enabling it to switch the optimal control strategy path in time according to changes in equipment operating status and environmental parameters, maintain the stability of the system's computing load under high-frequency data input, and reduce the complexity of the real-time control algorithm.

[0059] The present invention extracts the path with the highest fitting degree through the pattern fitting index set, and determines the target path direction in combination with the operating mode identifier. By calculating the difference vector and the state offset direction, it can dynamically analyze the deviation trend between the current input characteristics and the target control state, and then output a structured state transformation index. It can complete path fitting and state mapping within milliseconds, greatly improving the response speed and accuracy of path changes. By extracting the difference trend between the path characteristics and the input response in real time, it avoids control anomalies caused by state mismatch.

[0060] The present invention realizes the virtualization and serialized expression of physical circuit logic by mapping state transformation indexes into digital control sequences, and then constructs a structured control factor matrix, so that all logical changes are converted into digital signal operations. It breaks through the strong dependence of traditional control systems on hardware structure, greatly improves the flexibility of system structure and the dynamic expansion capability of control strategy, and can combine and encode logical tuple signals and control path numbers to construct a control factor matrix in a unified format, providing a basis for generating high-precision and high-consistency execution instructions.

[0061] The present invention generates execution instructions through a digital control matrix, has dynamic orchestration capabilities, and can adapt to different operating modes and circuit structures. Its control logic is completely generated by the upper-level module and is interpretable, which helps to achieve operation log tracking, output behavior verification and automatic fault location, thereby improving system operation and maintenance efficiency and safety level. At the same time, hierarchical control or hierarchical protection logic configuration is performed based on status instructions, and step-by-step response control measures are taken in a timely manner in key application scenarios to ensure stable system operation and prevent equipment damage. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a magnet power supply control system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0064] like Figure 1 As shown, an embodiment of the present invention provides a magnet power supply control system, the system comprising:

[0065] An input module is used to obtain parameter setting instructions and operation mode identifiers input by the control terminal, and extract the set current value, actual voltage value and historical operation sequence index value;

[0066] An aggregation module is used to determine the control input parameters at each time point based on the set current value and the actual voltage value, and construct a first data matrix; construct a second data matrix based on the historical operation sequence index value, and superimpose the first data matrix and the second data matrix to obtain a fused data tensor;

[0067] The fitting module is used to extract gradient features based on the fused data tensor, obtain a feature offset vector group, and calculate the response fit between it and each control path to obtain a pattern fitting index set;

[0068] A conversion module is used to extract the maximum response fit and determine the state conversion direction according to the operation mode identifier and the mode fitting index set, thereby obtaining a state conversion index;

[0069] A mapping module is used to call the control parameter mapping path according to the state conversion index, map the logical conditions of the physical circuit change into a digital control sequence, and obtain the control factor matrix;

[0070] The control module is used to generate execution instructions according to the control factor matrix and send them to the power output execution layer.

[0071] In an embodiment of the present invention, an input module is used to obtain parameter setting instructions and operation mode identifiers input by a control terminal, and to extract set current values, actual voltage values, and historical operation sequence index values, thereby ensuring that the input basis for system operation has a unified structure and multi-source data fusion capability, laying the foundation for subsequent aggregation analysis; an aggregation module is used to determine the control input parameters at each time point based on the set current values ​​and actual voltage values, and construct a first data matrix; a second data matrix is ​​constructed based on the historical operation sequence index values, and the first data matrix and the second data matrix are superimposed to obtain a fused data tensor, thereby ensuring that the fused data tensor has the structural characteristics of both timing information and configuration logic, thereby improving the system's contextual cognitive capability for state judgment, and making the control system no longer a static logic mapping, but a dynamic control body with historical response memory.

[0072] The fitting module extracts gradient features from the fused data tensor, generates a set of feature offset vectors, and calculates the response fit between these and each control path, resulting in a set of mode fit indices. By using the quantified response fit as a path optimization criterion, the system improves the objectivity and interpretability of control path selection. The transformation module extracts the maximum response fit based on the operating mode identifier and the set of mode fit indices, determines the state transition direction, and obtains a state transition index, thus avoiding the risk of relying solely on historical data while ignoring input fluctuations. By guiding state direction identification using difference vectors, the system possesses the ability to automatically correct for input disturbances, ensuring the controllability and safety of the control strategy switching process. The mapping module invokes the control parameter mapping path based on the state transition index, mapping the logical conditions of the physical circuit changes into digital control sequences, resulting in a control factor matrix. This enables the digital transfer of control strategies from the logic layer to the execution layer, effectively avoiding the response lag caused by fixed physical connections in traditional circuit control. The control module generates execution instructions based on the control factor matrix and sends them to the power output execution layer, ensuring real-time synchronization of the execution logic and enabling modular distribution to support complex multi-level control architectures.

[0073] The input module is used to obtain the parameter setting instructions and operation mode identification input by the control terminal, and extract the set current value, actual voltage value and historical operation sequence index value, specifically including:

[0074] First, a communication link is established with the control terminal. The communication link can be constructed through an Ethernet interface, a CAN bus, or a serial communication method to ensure that the parameter information is transmitted from the control terminal to the control system in real time. After the communication link is established, the input module receives the operation instruction set by the operator through a graphical interface or a script control interface. The instruction contains at least three key fields: a parameter setting instruction, an operation mode identifier, and a historical operation sequence call identifier. The set current value embedded in the parameter setting instruction is obtained by parsing the fixed byte position of the command packet. For example, the floating point value of the set current is stored in the 6th to 9th bytes in the hexadecimal data packet. After the input module parses it, it is stored in the structure Buffer; the actual voltage value is synchronously sampled by an external voltage sensor and input into the input module. The sampled value is processed by A / D conversion and recorded at the same timestamp as the set current value; the historical operation sequence index value is provided by the operation history database built into the system. After receiving the historical index call instruction, the input module extracts the corresponding index number according to the instruction field and passes the historical operation configuration vector pointed to by the index number to the subsequent module; during the entire data extraction process, the input module performs a one-time verification process on all data, including CRC verification, data type verification and field integrity verification. After the verification passes, the data is formatted into a standard input feature vector and packaged and transmitted to the aggregation module.

[0075] The control module is used to generate execution instructions based on the control factor matrix and send them to the power output execution layer, specifically including:

[0076] Receive the control factor matrix, which is in the form of a two-dimensional array, where each row corresponds to a control instruction logic block, and each column corresponds to the control dimension parameter under the logic block; first, perform instruction parsing on the control factor matrix, and the parsing rules are determined by the control instruction template file built into the system. For example, when the value of the first column of the control factor matrix is ​​"1", it indicates that the main power switch is triggered, the value of the second column is a numerical control parameter used to set the power output current level, and the third column is used to set the soft start time window, etc.; after the parsing is completed, the control module fills each control factor into the preset control instruction structure to construct a complete control command sequence; thereafter, the control module performs communication encapsulation on the control command to construct a data frame that complies with the power execution layer protocol , for example, using Modbus, CANopen or private protocol formats, and appending checksum fields and timestamp information; after encapsulation is completed, the control module sends the command to the power execution layer through the electrical interface via the underlying driver. Common interfaces include analog signal channels (for fast voltage / current control), digital IO channels (for logic switching control), PWM control ports (for output modulation), and high-bandwidth serial interfaces (such as SPI / RS485); the control module also has an instruction feedback mechanism. After sending the instruction, it starts the waiting response state and receives the execution result code and status identifier returned by the power output layer. Based on the feedback, it determines whether the instruction execution is successful. If not, it initiates the instruction resend mechanism and records the fault information.

[0077] In a preferred embodiment of the present invention, the aggregation module includes:

[0078] A sampling unit is used to extract control input parameters according to corresponding combinations of set current values ​​and actual voltage values ​​at each time point to obtain an input feature vector group;

[0079] An input matrix construction unit is used to arrange the input feature vector group into a two-dimensional array in time sequence, and use the input feature vectors as columns and the time points as rows to obtain a first data matrix;

[0080] A sequence mapping unit is used to extract control configuration parameters according to the historical operation sequence index value to obtain a historical configuration vector group;

[0081] The sequence matrix construction unit is used to arrange the historical configuration vector group into a two-dimensional array in index order, and use the historical configuration vectors as columns and the historical operation sequences as rows to obtain a second data matrix.

[0082] In an embodiment of the present invention, a sampling unit is used to extract control input parameters based on corresponding combinations of set current values ​​and actual voltage values ​​at various time points to obtain an input feature vector group, accurately capturing instantaneous deviations between control targets and actual operating results, and laying a stable foundation for subsequent data matrix construction. An input matrix construction unit is used to arrange the input feature vector group in chronological order into a two-dimensional array, using the input feature vectors as columns and the time points as rows, to obtain a first data matrix. This matrix can systematically record the changing patterns of the control input over time, allowing short-term input trends and mutation characteristics to be expressed in a structured form, providing a basis for subsequent pattern analysis and behavior fitting. A sequence mapping unit is used to extract control configuration parameters based on historical operation sequence index values ​​to obtain a historical configuration vector group, thereby achieving a structured restoration of historical control behaviors and enabling the system to utilize existing control mode experience to participate in current strategy generation and provide control behavior samples. A sequence matrix construction unit is used to arrange the historical configuration vector group in index order into a two-dimensional array, using the historical configuration vectors as columns and the historical operation sequences as rows, to obtain a second data matrix, completing the data structure reconstruction of historical control experience and providing a quantifiable historical control mapping method that can participate in algorithm matching.

[0083] The sampling unit is used to extract the control input parameters according to the corresponding combination of the set current value and the actual voltage value at each time point to obtain the input feature vector group, which specifically includes:

[0084] First, a synchronous sampling time reference is established, and a constant sampling period is set, such as sampling once every 10ms, to ensure data synchronization between the set current value and the actual voltage value. At each sampling time point, the system reads the excitation current value set at the current moment through the control terminal, and at the same time calls the sensor to collect the actual voltage value fed back in the current loop. In order to reduce the interference of data anomalies, a limiting mechanism is also set to discard or truncate data that exceeds the safety threshold of electrical parameters, and the exponential sliding average method is used to denoise and smooth the data at adjacent time points to improve the stability of the feature vector. Subsequently, the set current value and the actual voltage value corresponding to each sampling time point form a feature pair and are encoded as a control input parameter vector in a fixed order. This process is continuously executed throughout the sampling period, and eventually a group of input feature vectors arranged in time is formed.

[0085] The sequence mapping unit is used to extract the control configuration parameters according to the historical operation sequence index value to obtain the historical configuration vector group, which specifically includes:

[0086] First, the system's built-in historical configuration parameter database is accessed through a search mechanism. The database maintains data entries for multiple historical control modes by index number. Each record contains the corresponding configuration parameters such as the set current range, voltage control limit, power switch logic, magnetic field gain correction factor, cooling response time window, and temperature safety factor. The sequence mapping unit reads each index in sequence, locates the corresponding control parameter entry, and converts it into a vector form with a unified structure, namely the historical configuration vector. To ensure the comparability of the configuration data, the unit normalizes and standardizes the vector dimensions of all extracted configuration parameters, scales the parameters of each dimension to the same numerical range, and repairs missing fields by filling in the nearest value or replacing the most frequent historical value. After processing is completed, all historical configuration vectors will be arranged into a group of vector structures in index order. This group of structures is the historical configuration vector.

[0087] In a preferred embodiment of the present invention, the aggregation module further includes:

[0088] A sample matching unit is used to extract each column of input feature vectors from the first data matrix and each column of historical configuration vectors from the second data matrix to obtain an input configuration sample pair set;

[0089] A matching fusion scoring unit is used to match each pair of vectors in the input configuration sample pair set and calculate the matching fusion score of the sample pair;

[0090] The matching matrix unit is used to arrange the matching fusion scores according to the input feature vector number and the historical configuration vector number to obtain a two-dimensional matching matrix set, and use the two-dimensional matching matrix at each time point as an independent matching channel to obtain a three-dimensional fusion structure;

[0091] The fusion tensor construction unit is used to stack all independent matching channels in the three-dimensional fusion structure in chronological order, determine the fusion score of the input feature vector and the historical configuration group at the corresponding time point, and obtain the fusion data tensor.

[0092] In an embodiment of the present invention, a sample matching unit is used to extract each column of input feature vectors from the first data matrix and each column of historical configuration vectors from the second data matrix to obtain an input configuration sample pair set, establish a mapping relationship between the current state of the input and the historical control record, avoid scoring deviation or misjudgment of response path selection due to omission of matching samples, and lay a foundation for subsequent similarity evaluation; a matching fusion scoring unit is used to match each pair of vectors in the input configuration sample pair set, calculate the matching fusion score of the sample pair, capture the nonlinear similarity relationship between the input feature and the historical configuration, and make a unified comparison under different input scales to provide a reliable basis for the subsequent construction of the matching matrix; a matching matrix unit is used to match the matching fusion The scores are arranged according to the input feature vector number and the historical configuration vector number to obtain a two-dimensional matching matrix set, and the two-dimensional matching matrix at each time point is used as an independent matching channel to obtain a three-dimensional fusion structure. The static matching results are extended to the time series dimension, so that the matching relationship has continuity and evolution in time, which is helpful to observe the dynamic trajectory of similarity of input features as they change over time; the fusion tensor construction unit is used to stack all independent matching channels in the three-dimensional fusion structure in chronological order, determine the fusion score of the input feature vector and the historical configuration group at the corresponding time point, and obtain the fusion data tensor, realizing the spatial expansion of data from two dimensions to three dimensions, more accurately locating the optimal response path, and providing a basis for the dynamic adjustment of the control strategy.

[0093] The matching matrix unit is used to arrange the matching fusion scores according to the input feature vector number and the historical configuration vector number to obtain a two-dimensional matching matrix set, and use the two-dimensional matching matrices at each time point as independent matching channels to obtain a three-dimensional fusion structure, which specifically includes:

[0094] First, the system maps the matching fusion score between any input feature vector and any historical configuration vector to the element value in the matrix according to the number index of the input feature vector group and the number index of the historical configuration vector group, forming a two-dimensional matrix. This process is repeated once at each time point, that is, the scoring matrix construction process is independently performed on the input feature vector and historical configuration vector at each moment, and finally a set of two-dimensional matching matrices at different time points is obtained. The system then marks these two-dimensional matching matrices as independent channels on the time axis, that is, each matrix is ​​regarded as a scoring channel representing each time point, without numerical fusion, and this set of two-dimensional matching matrices is combined according to the time dimension to form a three-dimensional fusion structure tensor. ,in represents the input feature vector index dimension, Represents the historical configuration vector index dimension, Represents the time channel dimension.

[0095] The fusion tensor construction unit is used to stack all independent matching channels in the 3D fusion structure in chronological order, determine the fusion score of the input feature vector and the historical configuration group at the corresponding time point, and obtain the fusion data tensor, which specifically includes:

[0096] The system traverses all time channels in the 3D fusion structure , the elements in each two-dimensional matching matrix are used as the positioning parameters of the third dimension of the tensor based on their position in the time axis, and each matching score value is assigned to the element of the fusion tensor accordingly. In this process, in order to ensure the continuity of the tensor structure and the stability of the matching value, the system will normalize or interpolate all the score values ​​to ensure that the tensor remains intact at certain input missing time points. To avoid sudden changes in the score matrix at the end of the time period, the system can use a sliding window weighted average method to smooth the boundaries. The final fusion data tensor has a complete three-dimensional structure of input dimension-configuration dimension-time dimension, which can accurately express the matching strength relationship between each input feature and the historical configuration at each time point.

[0097] In a preferred embodiment of the present invention, the matching fusion scoring unit includes:

[0098] The matching fusion score calculation unit is used to perform feature mapping on the input feature vector to obtain a feature mapping value; and to perform normalization processing on the historical configuration vector to obtain a normalized strength value;

[0099] According to the feature map value and the normalized intensity value, the difference is calculated and nonlinearly amplified to obtain the fused difference term; the fused difference terms of all input feature vectors and historical configuration vectors are double-summed to obtain the total fused difference term;

[0100] The configuration response accumulation item is calculated based on the product of the historical configuration vector and the normalized intensity value; the total fusion difference item and the configuration response accumulation item are fused to obtain the fusion matching factor; the fusion matching factor is normalized to obtain the matching fusion score.

[0101] In an embodiment of the present invention, a matching fusion score calculation unit is used to perform feature mapping on an input feature vector to obtain a feature mapping value, thereby improving the expressive power of the input feature in a high-dimensional space and making the nonlinear correlation between features explicit, thereby providing a richer information basis for subsequent comparison with historical vectors; normalizing the historical configuration vector to obtain a normalized strength value, thereby eliminating matching deviations caused by inconsistent dimensions or differences in value distribution between different historical configuration vectors; calculating the difference between the feature mapping value and the normalized strength value and performing nonlinear amplification to obtain a fusion difference term; and performing a double summation on the fusion difference terms of all input feature vectors and historical configuration vectors to obtain a total fusion difference. The difference degree index between the input feature and the historical response is constructed by multiplying the historical configuration vector and the normalized intensity value, and the configuration response accumulation term is calculated, which reflects the stable response intensity of the historical configuration in its own space, avoiding relying solely on the difference value and ignoring the dominant role of the configuration itself on the output response; the total fusion difference term and the configuration response accumulation term are fused to obtain the fusion matching factor; the fusion matching factor is normalized to obtain the matching fusion score, which realizes the fusion judgment of the matching degree from the two directions of difference expression and response accumulation, reflecting both the similarity between the current input and historical experience and the explanatory power of historical samples for the current input, providing a basis for subsequent fitting path screening and state direction discrimination.

[0102] The calculation formula of the matching fusion score is:

[0103] ,

[0104] in, Score the matching fusion, is the index of the input feature vector, is the total number of input eigenvectors in the first data matrix, The index of the history configuration vector, is the total number of historical configuration vectors in the second data matrix, = , For the input feature vectors, = , For the History configuration vectors, , , and is the coefficient.

[0105] Among them, the coefficient Appears in the nonlinear difference term of the scoring formula numerator, in the form of , which is used to characterize the degree of difference between the input features and the historical configuration in the mapping space. It determines the degree of amplification of the system's response to the difference. If it is set high, the ability to identify small differences in features is enhanced, which is suitable for control environments that require high matching accuracy and low data noise, such as high-precision power output systems. If it is set too low, the system's response to feature differences is compressed, and the matching score is more tolerant to small deviations, which is suitable for actual engineering data with disturbances or measurement errors. The recommended value range is usually between 1.0 and 2.5. The specific value can be adjusted according to the range of normalized feature variance after data preprocessing. If the normalized feature difference is generally less than 0.2, it is recommended to increase it appropriately. , for example, set it to 2.0 to avoid the convergence of scoring results. If the difference fluctuates widely, it can be appropriately reduced to maintain the smoothness of the scoring. In specific engineering implementation, the cross-validation method or the objective function fitting error minimization strategy can be used to For optimization, you can also combine the distribution density curve of the sample scoring results to make corrections by minimizing the scoring kurtosis or improving the scoring range control ability.

[0106] coefficient It controls the nonlinear growth rate of the difference term, that is, the penalty of the scoring function on the matching error. Its role is to increase the original difference value to the power , thereby achieving nonlinear suppression or emphasis on samples with large differences. When the score function is 0, it grows linearly. At this time, the system maintains the same sensitivity to all differences. If it is greater than 0, it means that the system gives a greater degree of weight gain or penalty to samples with larger matching differences, thereby improving the score differentiation between high matching samples and medium and low matching samples. The value of is usually recommended to be controlled in the range of 0.2 to 0.6. When the control target tends to be conservative matching, that is, when the system is expected to make clear judgments on small changes between the input state and historical experience, it is advisable to It is recommended to set the value to 0.5 or higher to enhance the ability to suppress samples with large difference values. On the contrary, in scenarios where high fault tolerance is required or the quality of historical data fluctuates, it is recommended to lower the value appropriately. , for example, set it to 0.2 to avoid the scoring function being too sensitive.

[0107] In a preferred embodiment of the present invention, the fitting module includes:

[0108] A gradient extraction unit is used to calculate the response change rate of the input feature vector under different historical configuration vectors based on the fused data tensor to obtain a gradient tensor set;

[0109] The offset vector unit is used to compress the dimension of the gradient tensor set, aggregate it along the time axis, calculate the response equilibrium offset of each input feature vector under all historical configuration vectors, and obtain the feature offset vector group;

[0110] A control path mapping unit is used to determine the characteristic response structure of the input characteristics corresponding to different control logics according to the control path, and obtain a path feature vector set;

[0111] The fitting index unit is used to map the characteristic offset vector group and the path characteristic vector set, calculate the response fitting degree, and obtain the pattern fitting index set.

[0112] In an embodiment of the present invention, a gradient extraction unit is used to calculate the response change rate of the input feature vector under different historical configuration vectors based on the fused data tensor, and obtain a gradient tensor set, which accurately reflects the dynamic coupling relationship between the input feature and the control response, and provides a sufficient basis for subsequent path fitting; an offset vector unit is used to compress the dimension of the gradient tensor set, aggregate it along the time axis, calculate the response equilibrium offset of each input feature vector under all historical configuration vectors, and obtain a feature offset vector group, which quantifies the average response offset of the input feature in all historical configurations, avoids the problem of path judgment error caused by single abnormal data, and provides a basis for path matching. Direction reference in high-dimensional space; a control path mapping unit, which is used to determine the characteristic response structure of the input characteristics corresponding to different control logics according to the control path, obtain the path characteristic vector set, and deeply match the input characteristic response trend with the historical or preset control path to achieve accurate identification of the control behavior and judgment of the path directionality; a fitting index unit, which is used to map the characteristic offset vector group and the path characteristic vector set, calculate the response fitting degree, and obtain the pattern fitting index set, thereby realizing the quantitative evaluation of the control path in the input response dimension, so that the selection of the control strategy has an objective mathematical basis, and avoids the subjective judgment of the traditional system based on artificial weights or experience optimization.

[0113] The gradient extraction unit is used to calculate the response change rate of the input feature vector under different historical configuration vectors based on the fused data tensor to obtain a gradient tensor set, which specifically includes:

[0114] The system receives a fused data tensor containing the numerical changes of the input feature vector at multiple time points and under multiple historical configuration vectors. This fused data tensor is a three-dimensional tensor structure, with dimensions corresponding to the time series, the input feature dimension, and the historical configuration dimension. To extract the response change rate of each input feature vector under different historical configuration conditions, the system first slices the data according to the input feature dimension, extracting the two-dimensional time-configuration submatrix corresponding to each dimension in the fused data tensor. Each submatrix is ​​then differentiated along the time axis to calculate the incremental change of adjacent input feature values ​​under different configuration conditions at each time point. Smoothing is performed using methods such as central differencing or sliding weighted differencing to reduce noise interference caused by sensor sampling fluctuations. Based on the difference results, the system further normalizes each historical configuration dimension, mapping the change rate to the interval [-1, 1] to unify the response dimensions and amplitudes of different input features. Finally, the response change rates corresponding to all input feature dimensions are reassembled into a three-dimensional array structure, forming a gradient tensor set, where each element represents the response slope or trend direction of a specific input feature under a specific configuration dimension and corresponding time point.

[0115] The control path mapping unit is used to determine the characteristic response structure of the input features corresponding to different control logics according to the control path, and obtain a path feature vector set, which specifically includes:

[0116] The control path set is preset by the system or generated during the training phase and contains multiple control paths. Each control path consists of a set of path feature vectors that describe the desired input feature response pattern under a specific control strategy. First, the system sequentially reads the path feature vectors of each control path from the path set, maps them as vectors to the same dimensional space as the input feature vectors, and normalizes them to unify the amplitude dimensions. The system then structurally aligns the feature offset vector set with the path feature vectors of each control path, establishing a dimensional correspondence between the input response and the target path structure. Based on this, the system calculates the directional consistency coefficient between each pair of feature offset vectors and path feature vectors. The degree of match can be assessed using methods such as angle cosine, projected distance, or relative entropy. For each control path, the system aggregates the directional matching results across all input dimensions to generate a complete path matching mapping table. Using this mapping table, the system accurately identifies the path structure that best matches the current input response offset trend and obtains the path feature vector set.

[0117] In a preferred embodiment of the present invention, the offset vector unit includes:

[0118] The temporal aggregation unit is used to average the gradient values ​​of the same input feature dimension at all time points in the gradient tensor set to obtain the response mean matrix;

[0119] The configuration alignment unit is used to perform structural matching on each row in the response mean matrix and the historical configuration vector number, and uniformly map them to the corresponding historical configuration dimension to obtain the configuration aggregation matrix;

[0120] An offset calculation unit is used to perform standard deviation compression and normalization on the response mean of each input feature dimension in the configuration aggregation matrix to obtain a response equalization offset value;

[0121] The vector generation unit is used to arrange the response equalization offset values ​​of all input feature dimensions into a vector structure in dimensional order to obtain a feature offset vector group.

[0122] In an embodiment of the present invention, a time aggregation unit is used to average the gradient values ​​of the same input feature dimension in the gradient tensor set at all time points to obtain a response mean matrix, which can effectively alleviate data noise interference caused by sampling fluctuations or transient anomalies. The configuration alignment unit is used to structurally match each row in the response mean matrix with the historical configuration vector number and uniformly map it to the corresponding historical configuration dimension to obtain a configuration aggregation matrix, thereby achieving consistent conversion of the data structure and avoiding calculation errors or vector mismatch problems caused by inconsistent dimensions in subsequent offset calculations. The offset calculation unit is used to perform standard deviation compression and normalization on the response mean of each input feature dimension in the configuration aggregation matrix to obtain a response balanced offset value. The standard deviation compression operation on the response mean can effectively eliminate the impact of extreme values ​​on the offset result, and the normalization process ensures that different input features are comparable in numerical range. The vector generation unit is used to arrange the response balanced offset values ​​of all input feature dimensions into a vector structure in dimensional order to obtain a feature offset vector group, thereby achieving a structural transformation from local response statistics to the overall input feature expression and providing unified vector space support for pattern fitting.

[0123] The configuration alignment unit is used to structurally match each row in the response mean matrix with the historical configuration vector number and uniformly map it to the corresponding historical configuration dimension to obtain a configuration aggregation matrix, which specifically includes:

[0124] First, the response mean matrix is ​​received. This matrix is ​​a two-dimensional structure, with rows corresponding to the input feature dimensions and columns representing the numbers of different historical configuration vectors. Because the parameter distribution and field structure of each historical configuration vector may differ in the control strategy, directly using this matrix for offset feature analysis will face dimensional inconsistency. To achieve consistent data structure matching, the system first calls a pre-established historical configuration vector number mapping table, which defines the standard parameter structure and field correspondence for each historical configuration. For each row of the response mean matrix corresponding to the input feature dimension, the historical configuration number information in the column is traversed and the corresponding standard configuration parameter structure is retrieved from the mapping table. The response mean value items are compared with the field order and attribute type in the target structure to perform corresponding field-level rearrangement and interpolation. If some historical configurations lack fields in the current standard structure, they are supplemented by stretching the previous and next time periods or adjacent configuration data using linear interpolation or sample averaging. Finally, all historical configuration vectors are converted to a unified standard structure, and the output is a configuration aggregation matrix.

[0125] The offset calculation unit is used to perform standard deviation compression and normalization on the response mean of each input feature dimension in the configuration aggregation matrix to obtain a response equalization offset value, specifically including:

[0126] First, each row of data in the configuration aggregation matrix is ​​processed. For any row of input feature dimensions, it corresponds to the response mean under multiple historical configuration dimensions, which is recorded as a set of sequences. The mean and standard deviation of this sequence are calculated, where the mean reflects the average response level of the input feature in the historical configuration, and the standard deviation measures its response volatility under different configuration conditions. The response mean sequence is then subjected to standard deviation compression to prevent individual outliers from having a dominant influence on the overall offset estimation result. The compression operation usually normalizes the standard deviation by adjusting it, for example, comparing it with the system's preset minimum variance threshold, and then dividing the response mean by the standard deviation compression result to obtain the final normalized response offset value. This response balance offset value can be considered as a comprehensive characterization of the average response trend and response stability of the input feature dimension under all historical configurations, with the advantages of dimension elimination and scale unification. Finally, the response balance offset values ​​of all input feature dimensions are spliced ​​into a vector form in dimensional order.

[0127] In a preferred embodiment of the present invention, the fitting index unit includes:

[0128] A response fitting calculation unit is used to calculate the characteristic mean of the path characteristic vector under different control paths;

[0129] Based on the offset eigenvector and the path eigenvector of the input eigenvector, the numerical projection ratio of the two on each input feature dimension is calculated to obtain a weighted projection term; based on the absolute difference between the offset eigenvector and the path eigenvector, a symmetric normalized ratio is constructed to obtain a nonlinear suppression term; the weighted projection term and the nonlinear suppression term are multiplied item by item according to the dimension number and summed to obtain the total path fitting term;

[0130] The variance of the path eigenvector is calculated, and the stability penalty term is constructed based on it and the eigenmean. The total path fitting term is used as the numerator and the stability penalty term as the denominator to obtain the response fitting degree.

[0131] In an embodiment of the present invention, a response fitting calculation unit is used to calculate the characteristic mean of the path characteristic vector under different control paths, effectively extract the central trend of the response characteristics of each control path, and provide a reference standard for subsequent input state matching; according to the offset characteristic vector and the path characteristic vector of the input characteristic vector, the numerical projection ratio of the two on each input characteristic dimension is calculated to obtain a weighted projection term, quantify the projection effect of the input state on the path response characteristics, identify the control path closest to the current characteristic offset direction, and ensure that the subsequently selected path has a high degree of physical feasibility and strategy continuity; according to the absolute difference between the offset characteristic vector and the path characteristic vector, a symmetric normalized ratio is constructed to obtain a nonlinear suppression term, effectively control the fitting error generated when there is a large offset between the input state and the path response, and ensure the uniformity of vector difference processing on the numerical scale. The weighted projection term and the nonlinear suppression term are multiplied and summed up item by item according to the dimension number to obtain the total path fitting term, which effectively improves the system's sensitivity to the matching accuracy of key dimensions, ensures that the scoring results not only reflect the average level, but also focus on the response consistency in key dimensions, making the path matching results more discriminative and physically meaningful; calculates the variance of the path eigenvector, and constructs a stability penalty term based on its sum and the characteristic mean to suppress the interference of the unstable response path on the final fitting result, avoid selecting the control path that is susceptible to disturbance or has unstable historical response, and ensure the controllability and predictability of the path selection; takes the total path fitting term as the numerator and the stability penalty term as the denominator to obtain the response fitting degree, realizing an adaptive scoring mechanism under multiple paths, taking into account the two core indicators of fitting matching and path stability, and providing a basis for dynamic path decision-making and the realization of intelligent power supply control.

[0132] The calculation formula of the response fitting degree is:

[0133] ,

[0134] in, is the response fit, For the The offset feature vector of the input feature vector, For the Under the control path The path feature vector of the input feature vectors, To control the index of the path, For the The characteristic mean of all path characteristic vectors under the control path, is the variance of all path eigenvectors, is a constant, and is the coefficient.

[0135] in, is a minimum constant to avoid division by zero errors.

[0136] coefficient The role of is to regulate the suppression strength of the input features and path features in the nonlinear difference part. It is located in the power exponent in the numerator and is used to strengthen or weaken the influence of the difference between the two vectors on the overall fit. If the system needs to more strongly suppress the local deviation between the input state and the path response (for example, to deal with high-frequency noise disturbance scenarios), it should be appropriately increased. On the contrary, in a scenario where a certain offset range is allowed and the control system focuses more on overall trend fitting rather than detail consistency, Can be reduced moderately. The value range is 1.5 to 3.0. The specific value can be adjusted by experimentally verifying the sensitivity of the path response to disturbances. For example, in a magnet control system, if the stability of the path switching is prioritized rather than the sensitivity, Set to 2.5 to suppress the excessive amplification of the response fitting degree to the single point difference. If you want the system to have a strong adaptive response capability, such as to cope with the dynamic cooling water flow change scenario, you can set Set it to around 1.8.

[0137] coefficient The stability penalty term that appears in the denominator controls the weight of the path feature vector amplitude on the fit. When a path feature vector has a local deviation but the overall mean remains consistent, whether it is necessary to penalize its possibility as a path selection candidate. If the system has high requirements for path stability, that is, the characteristics of the control path cannot change too much, then it should be increased. to increase the severity of the penalty. The value range is 0.5 to 2.0. Its setting depends on the standard deviation of the path characteristic distribution and the tolerance of the control accuracy. For example, when the control path response fluctuation range in the magnet power system is large, or there are obvious differences between different historical paths, it is recommended to The value is set to 1.5 to enhance the ability to filter out unstable paths. In a system where the path response fluctuation is relatively small, if more response paths need to be retained as candidate solutions, Set it below 0.8 to improve system fault tolerance and flexibility.

[0138] In a preferred embodiment of the present invention, the conversion module includes:

[0139] The matching extraction unit is used to traverse and compare all response fitting degrees in the pattern fitting index set, extract the control path number with the largest fitting degree value, and obtain the maximum fitting path index;

[0140] The mode recognition unit is used to filter out the control path index corresponding to the current operation mode according to the operation mode identifier, and compare it with the maximum fitting path index to obtain the target path number;

[0141] The state direction unit is used to extract the corresponding path feature vector according to the target path number, and perform difference calculation with the feature offset vector of the corresponding dimension to obtain a response difference vector. Then, based on the response difference vector, the adjustment direction of the current input feature is determined to obtain the state offset direction;

[0142] The index generating unit is used to encode and map the maximum fitting path index with the state offset direction to obtain the state conversion index.

[0143] In an embodiment of the present invention, a matching extraction unit is used to traverse and compare all response fits in a pattern fitting index set, extract the control path number with the largest fit value, obtain the maximum fit path index, and compare the response fits of all paths to ensure that the system selects the control logic path that best matches the current operating state; a pattern recognition unit is used to screen out the control path index corresponding to the current operating mode according to the operating mode identifier, and compare it with the maximum fit path index to obtain the target path number, thereby realizing a consistency verification mechanism for the operating strategy and ensuring the logical correctness and strategy consistency of the path selection; a state direction unit is used to select the control path index corresponding to the current operating mode according to the target path number, The corresponding path feature vector is extracted and the difference is calculated with the feature offset vector of the corresponding dimension to obtain the response difference vector. Then, based on the response difference vector, the adjustment direction of the current input feature is determined to obtain the state offset direction, which effectively constructs the offset criterion basis of the input state, enhances the system's ability to recognize input state trends and guide behavior, and provides a quantitative basis for control decisions; the index generation unit is used to encode and map the maximum fitting path index with the state offset direction to obtain the state transformation index, which effectively expresses the behavioral response logic of the system in the current state, ensures that the subsequent control factor calls have precise directionality, and realizes standardized switching and seamless control connection between different logical paths.

[0144] The mode recognition unit is used to filter out the control path index corresponding to the current operating mode according to the operating mode identifier, and compare it with the maximum fitting path index to obtain the target path number, specifically including:

[0145] The operating mode identifier generally represents the operating situation of the current system in the form of a scalar or string, such as predefined labels such as "steady-state operation", "temperature control protection", "abnormal cooling" or "energy consumption minimization mode". The mode recognition unit reads the operating mode identifier and uses it as a keyword to query the operating mode mapping table in the path configuration library. The mapping table is a pre-set structured data set that records each operating mode and its corresponding candidate control path number set. The path number is uniquely identified by an integer index. The system maps and searches the operating mode identifier through a hash map or a bidirectional index structure, and filters out a set of all path numbers associated with the mode, which is passed to the comparison process as a candidate path index set. In order to enhance the system's adaptability to non-standard modes or boundary states, the mapping table allows the configuration of a many-to-one mapping relationship, that is, multiple mode identifiers can correspond to the same path set, or a path number can belong to multiple mode sets at the same time, thereby enhancing the redundancy and fault tolerance of path identification.

[0146] After obtaining the candidate path index set, the system compares it with the maximum-fit path index output by the matching extraction unit. This comparison is typically performed using a set matching approach. If the maximum-fit path index is included in the candidate path index set, the system assumes that the current operating state matches the historically fitted path, and the target path number is the maximum-fit path index. If it is not included, the current operating mode places strong constraints on the path logic, and the system uses the preferred path number in the candidate path set as the alternative path output, forming the revised target path number.

[0147] In a preferred embodiment of the present invention, the state direction unit includes:

[0148] A feature extraction unit is used to extract the path feature vector with the corresponding number from the path feature vector set according to the target path number to obtain the target path vector;

[0149] a difference construction unit, configured to perform dimension-by-dimension difference operation on the target path vector and the offset vector of the corresponding dimension in the feature offset vector group to obtain a response difference vector;

[0150] A symbol parsing unit is used to parse the symbol information of each dimension in the response difference vector, mark the offset direction of the current input feature compared to the control, and obtain an adjustment direction set;

[0151] The trend aggregation unit is used to count the distribution trend of the overall offset direction according to the adjustment direction set, and determine a unified direction to obtain the state offset direction.

[0152] In an embodiment of the present invention, a feature extraction unit is used to extract the path feature vector of the corresponding number from the path feature vector set according to the target path number to obtain the target path vector, thereby ensuring that the system can accurately locate the most suitable control path feature structure under the current operating state and provide a direct reference benchmark for subsequent calculations; a difference construction unit is used to perform dimension-by-dimension difference operation on the target path vector and the offset vector of the corresponding dimension in the feature offset vector group to obtain a response difference vector, thereby quantifying the numerical difference between the current response state of the input feature and the expected control state; a symbol parsing unit is used to parse the symbol information of each dimension in the response difference vector, mark the offset direction of the current input feature compared to the control, obtain an adjustment direction set, map the original difference vector into a symbol vector, retain the direction information of the parameter offset, and avoid adjustment misjudgment due to numerical noise; a trend aggregation unit is used to count the distribution trend of the overall offset direction according to the adjustment direction set, determine a unified direction, obtain the state offset direction, integrate the multi-dimensional difference directions into a unified direction identifier, and effectively avoid adjustment conflicts caused by local dimensional differences.

[0153] The difference construction unit is used to perform dimension-by-dimension difference operation on the target path vector and the offset vector of the corresponding dimension in the feature offset vector group to obtain a response difference vector, specifically including:

[0154] After the system completes the extraction of the target path number, it obtains the target path vector corresponding to the number from the path feature vector set, and obtains the offset vector with the same dimensional structure from the feature offset vector group. The two constitute a set of vector pairs with structural consistency; then, the difference calculation is performed on each dimensional element in the target path vector and the element of the corresponding dimension in the offset vector in turn to obtain the response offset difference under that dimension, and then construct a complete response difference vector; to ensure the stability and analyzability of the difference result in subsequent steps, the system adopts the same numerical accuracy standard for all dimensional difference operations, and sets a dynamic truncation threshold during the calculation process to avoid the accumulation of very small errors causing numerical mutations; each dimensional element of the response difference vector represents the degree and directionality of the current input feature compared to the target path under that feature dimension.

[0155] The symbol parsing unit is used to parse the symbol information of each dimension in the response difference vector, mark the offset direction of the current input feature compared to the control, and obtain the adjustment direction set, which specifically includes:

[0156] The system reads each dimensional element in the response difference vector in turn and performs logical judgment on the sign information of each element; if the value of the dimensional element is greater than the preset positive offset threshold, the system determines it as a positive offset and marks it as +1; if the dimensional element is less than the set negative offset threshold, it is marked as -1; if the dimensional value is between the positive and negative offset thresholds, it is considered to have no significant offset and is marked as 0; the above operation is repeated for each dimension, and finally the symbol marking results of all dimensions are combined to form an adjustment direction set, each element in the direction set is a standardized symbol value, corresponding to the adjustment direction identifier of each dimension in the input feature vector; in order to avoid the direction judgment being affected by numerical noise, a symbol stabilization strategy can be introduced, such as using a sliding average method to smooth the response difference vector and then perform symbol judgment, thereby enhancing the robustness of direction resolution, and finally obtaining the adjustment direction set as the system's preliminary judgment result on the input feature adjustment trend.

[0157] The trend aggregation unit is used to calculate the distribution trend of the overall offset direction according to the adjustment direction set, determine a unified direction, and obtain the state offset direction, which specifically includes:

[0158] The system first groups and counts the elements in the adjustment direction set, and calculates the number of positive marks (+1), negative marks (-1) and zero marks (0) respectively to form the direction statistics results; after obtaining the statistical results, the system makes trend judgments based on the proportion of the number of each type of direction marks: if the number of positive marks is significantly greater than the number of negative marks and exceeds the set threshold ratio, the overall trend is judged to be a positive offset; conversely, if the number of negative marks is significantly dominant, it is judged to be a negative offset; if the number of positive and negative direction marks is equal or most of the dimensional marks are zero, the system determines the final direction based on the path stability criterion under the current operating mode or the offset trend of the last few cycles, ensuring that the system still has decision-making capabilities when the direction is unclear, and finally uses the determined unified offset direction as the state offset direction.

[0159] In a preferred embodiment of the present invention, the mapping module includes:

[0160] The path calling unit is used to match the state conversion index with the control parameter mapping path to obtain the parameter configuration path;

[0161] A conditional parsing unit is used to read the circuit logic change conditions in the parameter configuration path and parse the module change signals therein to obtain a logic tuple sequence;

[0162] a digital generation unit for combining and encoding the module change signal and the control path number in the logic tuple sequence to obtain a digital control sequence;

[0163] The factor matrix unit is used to expand the sequence data into a two-dimensional control matrix structure according to the digital control sequence to obtain a control factor matrix.

[0164] In an embodiment of the present invention, a path calling unit is used to match the control parameter mapping path according to the state conversion index to obtain the parameter configuration path, so as to realize the rapid retrieval and precise positioning of the corresponding control path by the system under complex control state, thereby avoiding the path mismatch problem caused by manual screening or reliance on preset branch judgment in the traditional system; a conditional parsing unit is used to read the circuit logic change condition in the parameter configuration path and parse the module change signal therein to obtain a logic tuple sequence, identify and structure the circuit change behavior required for the current control path, and solve the problems of logical confusion and parameter conflict when switching control modes in the traditional system; a digital generation unit is used to combine and encode the module change signal in the logic tuple sequence with the control path number to obtain a digital control sequence, map the complex logical behavior into a simple digital instruction sequence, and effectively reduce the parsing load of the system at the execution layer; a factor matrix unit is used to expand the sequence data into a two-dimensional control matrix structure according to the digital control sequence to obtain a control factor matrix, thereby realizing matrix management and modular control of logic signals at the execution layer.

[0165] The conditional parsing unit is used to read the circuit logic change conditions in the parameter configuration path and parse the module change signals therein to obtain a logic tuple sequence, specifically including:

[0166] First, the parameter configuration path corresponding to the index is extracted from the control parameter mapping path. This parameter configuration path stores multiple circuit logic change conditions, each of which defines its affected module, trigger threshold, and logic type in a structured manner. Using modules as indexes, the logic condition items within the path are scanned one by one. Each logic condition includes the target module number, change type (such as voltage increase, current cutoff, feedback enable), trigger method (such as duration, fluctuation threshold, state transition), and execution direction. The logic condition is parsed into a standardized module change signal, which expresses the specific action the control module must complete, such as "Module_05: SwitchToLow" or "Module_12: EnableByThreshold." A logic tuple structure is then constructed using triples, including fields (module number, change action, and condition type). All logic tuples are aggregated and arranged to generate a logic tuple sequence. The order of this sequence matches the node order in the configuration path to ensure data timing consistency.

[0167] The digital generation unit is used to combine and encode the module change signal and the control path number in the logic tuple sequence to obtain a digital control sequence, which specifically includes:

[0168] First, each module number is assigned a unique module identifier (e.g., an integer ID). A set of standard action codes is defined for each change action (e.g., 01 for voltage reduction, 02 for signal disconnection, 03 for feedback enable), and type identifiers are assigned for different condition types (e.g., A for threshold trigger, B for duration trigger). The encoding process uses the module identifier as the high-order segment, and the action code and condition identifier as the middle and low-order segments. Bit shifting or string concatenation constructs a uniformly formatted code segment. For example, a logic tuple (Module_08, SwitchToLow, ThresholdTrigger) might be encoded as 0801A, indicating that module 08 is executing action 01 with a trigger type of A. The system then prefixes each code segment with the current path number to form the complete control unit code. For example, for path number 21, the combined code is 210801A. All code segments are arranged in the original order of the logic tuples to form a one-dimensional array structure, ultimately generating a complete digital control sequence. This encoding method supports binary, decimal, or hexadecimal representation, depending on the platform.

[0169] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A magnet power supply control system, characterized in that: The system comprises: An input module is used to obtain parameter setting instructions and operation mode identifiers input by the control terminal, and extract the set current value, actual voltage value and historical operation sequence index value; An aggregation module is used to determine the control input parameters at each time point based on the set current value and the actual voltage value, and construct a first data matrix; construct a second data matrix based on the historical operation sequence index value, and superimpose the first data matrix and the second data matrix to obtain a fused data tensor; The fitting module is used to extract gradient features based on the fused data tensor, obtain a feature offset vector group, and calculate the response fit between it and each control path to obtain a pattern fitting index set; A conversion module is used to extract the maximum response fit and determine the state conversion direction according to the operation mode identifier and the mode fitting index set, thereby obtaining a state conversion index; A mapping module is used to call the control parameter mapping path according to the state conversion index, map the logical conditions of the physical circuit change into a digital control sequence, and obtain the control factor matrix; A control module is used to generate execution instructions according to the control factor matrix and send them to the power output execution layer; The aggregation module includes: A sampling unit is used to extract control input parameters according to corresponding combinations of set current values ​​and actual voltage values ​​at each time point to obtain an input feature vector group; An input matrix construction unit is used to arrange the input feature vector group into a two-dimensional array in time sequence, and use the input feature vectors as columns and the time points as rows to obtain a first data matrix; A sequence mapping unit is used to extract control configuration parameters according to the historical operation sequence index value to obtain a historical configuration vector group; The sequence matrix construction unit is used to arrange the historical configuration vector group into a two-dimensional array in index order, and use the historical configuration vectors as columns and the historical operation sequences as rows to obtain a second data matrix.

2. A magnet power supply control system according to claim 1, characterized in that: The aggregation module further includes: A sample matching unit is used to extract each column of input feature vectors from the first data matrix and each column of historical configuration vectors from the second data matrix to obtain an input configuration sample pair set; A matching fusion scoring unit is used to match each pair of vectors in the input configuration sample pair set and calculate the matching fusion score of the sample pair; The matching matrix unit is used to arrange the matching fusion scores according to the input feature vector number and the historical configuration vector number to obtain a two-dimensional matching matrix set, and use the two-dimensional matching matrix at each time point as an independent matching channel to obtain a three-dimensional fusion structure; The fusion tensor construction unit is used to stack all independent matching channels in the three-dimensional fusion structure in chronological order, determine the fusion score of the input feature vector and the historical configuration group at the corresponding time point, and obtain the fusion data tensor.

3. A magnet power supply control system according to claim 2, characterized in that: The matching fusion scoring unit includes: The matching fusion score calculation unit is used to perform feature mapping on the input feature vector to obtain a feature mapping value; and to perform normalization processing on the historical configuration vector to obtain a normalized strength value; According to the feature map value and the normalized intensity value, the difference is calculated and nonlinearly amplified to obtain the fused difference term; the fused difference terms of all input feature vectors and historical configuration vectors are double-summed to obtain the total fused difference term; The configuration response accumulation item is calculated based on the product of the historical configuration vector and the normalized intensity value; the total fusion difference item and the configuration response accumulation item are fused to obtain the fusion matching factor; the fusion matching factor is normalized to obtain the matching fusion score.

4. A magnet power supply control system according to claim 3, characterized in that: The fitting module includes: A gradient extraction unit is used to calculate the response change rate of the input feature vector under different historical configuration vectors based on the fused data tensor to obtain a gradient tensor set; The offset vector unit is used to compress the dimension of the gradient tensor set, aggregate it along the time axis, calculate the response equilibrium offset of each input feature vector under all historical configuration vectors, and obtain the feature offset vector group; A control path mapping unit is used to determine the characteristic response structure of the input characteristics corresponding to different control logics according to the control path, and obtain a path feature vector set; The fitting index unit is used to map the characteristic offset vector group and the path characteristic vector set, calculate the response fitting degree, and obtain the pattern fitting index set.

5. A magnet power supply control system according to claim 4, characterized in that: The offset vector unit includes: The temporal aggregation unit is used to average the gradient values ​​of the same input feature dimension at all time points in the gradient tensor set to obtain the response mean matrix; The configuration alignment unit is used to perform structural matching on each row in the response mean matrix and the historical configuration vector number, and uniformly map them to the corresponding historical configuration dimension to obtain the configuration aggregation matrix; An offset calculation unit is used to perform standard deviation compression and normalization on the response mean of each input feature dimension in the configuration aggregation matrix to obtain a response equalization offset value; The vector generation unit is used to arrange the response equalization offset values ​​of all input feature dimensions into a vector structure in dimensional order to obtain a feature offset vector group.

6. A magnet power supply control system according to claim 5, characterized in that: The fitting index unit includes: A response fitting calculation unit is used to calculate the characteristic mean of the path characteristic vector under different control paths; Based on the offset eigenvector and the path eigenvector of the input eigenvector, the numerical projection ratio of the two on each input feature dimension is calculated to obtain a weighted projection term; based on the absolute difference between the offset eigenvector and the path eigenvector, a symmetric normalized ratio is constructed to obtain a nonlinear suppression term; the weighted projection term and the nonlinear suppression term are multiplied item by item according to the dimension number and summed to obtain the total path fitting term; The variance of the path eigenvector is calculated, and the stability penalty term is constructed based on it and the eigenmean. The total path fitting term is used as the numerator and the stability penalty term as the denominator to obtain the response fitting degree.

7. A magnet power supply control system according to claim 6, characterized in that: The conversion module comprises: The matching extraction unit is used to traverse and compare all response fitting degrees in the pattern fitting index set, extract the control path number with the largest fitting degree value, and obtain the maximum fitting path index; The mode recognition unit is used to filter out the control path index corresponding to the current operation mode according to the operation mode identifier, and compare it with the maximum fitting path index to obtain the target path number; The state direction unit is used to extract the corresponding path feature vector according to the target path number, and perform difference calculation with the feature offset vector of the corresponding dimension to obtain a response difference vector. Then, based on the response difference vector, the adjustment direction of the current input feature is determined to obtain the state offset direction; The index generating unit is used to encode and map the maximum fitting path index with the state offset direction to obtain the state conversion index.

8. A magnet power supply control system according to claim 7, characterized in that: The state direction unit includes: A feature extraction unit is used to extract the path feature vector with the corresponding number from the path feature vector set according to the target path number to obtain the target path vector; a difference construction unit, configured to perform dimension-by-dimension difference operation on the target path vector and the offset vector of the corresponding dimension in the feature offset vector group to obtain a response difference vector; A symbol parsing unit is used to parse the symbol information of each dimension in the response difference vector, mark the offset direction of the current input feature compared to the control, and obtain an adjustment direction set; The trend aggregation unit is used to count the distribution trend of the overall offset direction according to the adjustment direction set, and determine a unified direction to obtain the state offset direction.

9. A magnet power supply control system according to claim 8, characterized in that: The mapping module includes: The path calling unit is used to match the state conversion index with the control parameter mapping path to obtain the parameter configuration path; A conditional parsing unit is used to read the circuit logic change conditions in the parameter configuration path and parse the module change signals therein to obtain a logic tuple sequence; a digital generation unit for combining and encoding the module change signal and the control path number in the logic tuple sequence to obtain a digital control sequence; The factor matrix unit is used to expand the sequence data into a two-dimensional control matrix structure according to the digital control sequence to obtain a control factor matrix.

Images