A multi-channel time-series data error bound quantization method and system based on dynamic graph grouping and error distribution

Abstract

The application relates to the technical field of industrial Internet of Things, and discloses a multi-channel time sequence data error bound quantification method and system based on dynamic graph grouping and error distribution, which comprises the following steps: S1: acquiring multi-channel time sequence data collected at the same collection point; S2: determining a scale factor, a point-by-point maximum absolute error upper bound, normalized channel time sequence data, and normalized error bounds of all channels; S3: calculating correlation measurement values of any two channels, and constructing a channel correlation graph; S4: determining a grouping threshold value, and forming an anchor point set from selected anchor point channels; S5: constructing a cross-channel correlation prediction model, and constructing an innovation sequence; S6: determining an error sensitivity coefficient, and determining a quantified error bound; S7: obtaining normalized reconstruction values of the anchor point channels and the innovation sequence; and S8: obtaining normalized reconstruction values of non-anchor point channels, and performing inverse normalization on the normalized reconstruction values according to the scale factor to obtain reconstruction values of the channels under the original dimension. The application can limit compress and highly restore original data.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) technology, specifically to a method and system for quantizing the error boundary of multi-channel time-series data based on dynamic graph grouping and error allocation. Background Technology

[0002] With the popularization of IoT and industrial digitalization technologies, more and more industrial gateways and smart terminals are beginning to continuously collect multi-channel time-series data (such as voltage, current, vibration, temperature, etc.) from the same equipment node (such as large generator sets and substation nodes) at high frequency. This type of multi-channel time-series data is not only massive in scale, but also often exhibits strong physical coupling and spatial correlation between channels. At the same time, affected by external factors such as equipment start-up and shutdown, operating condition switching, and environmental interference, multi-channel time-series data and its correlations show significant non-stationarity and dynamic evolution characteristics. In order to achieve real-time processing and uploading of this massive amount of data on edge devices with extremely limited transmission bandwidth and storage space, lossy compression with error limits has become a key means of cost reduction and efficiency improvement.

[0003] However, existing error limit compression methods generally treat each sensor channel as an independent single-channel sequence, completely ignoring the rich cross-channel spatial redundancy between physical quantities at the same acquisition point, resulting in a serious bottleneck in the overall compression ratio; even if some improved methods attempt to use the correlation between channels for cross-prediction, they often rely on fixed channel grouping or static reference strategies.

[0004] When faced with non-stationary industrial time-series data, this static strategy cannot adapt to the dynamic changes in channel correlation. More critically, since the reference channel itself contains quantization error, this error will be uncontrollably amplified and propagated in the cross-channel prediction model, causing the reconstructed data of the target channel to frequently exceed the strict error limit set by the user, thereby seriously undermining the reliability of high-precision tasks such as downstream fault diagnosis and condition monitoring.

[0005] Especially on edge acquisition nodes with limited computing power and storage space, existing methods struggle to simultaneously achieve both maximizing compression ratios through deep mining of multi-channel spatial joint redundancy and strictly blocking error propagation in complex cross-channel predictions to ensure absolute point-by-point error limits. Therefore, there is an urgent need in this field for a multi-channel data quantization method capable of sensing dynamic correlations between channels and adaptively adjusting groupings, while also possessing a scientific error budget allocation mechanism to curb error propagation. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, one of the objectives of this invention is to provide a multi-channel time series data error boundary quantization method based on dynamic graph grouping and error allocation. This method can solve the problem that existing multi-channel time series data quantization methods are unable to fully exploit the spatial correlation redundancy across channels, and that errors are easily amplified and propagated at each level when performing cross-channel prediction, causing the reconstructed data to exceed the strict error limit.

[0007] The technical solution adopted in this invention is as follows: A method for error boundary quantization of multi-channel time-series data based on dynamic graph grouping and error allocation, comprising the following steps:

[0008] S1: Acquire multi-channel time-series data collected from the same acquisition point;

[0009] S2: Determine the scaling factor for each channel and the upper bound of the maximum absolute error of each channel under the original dimensions. Obtain the normalized channel time series data through the scaling factor of each channel and the corresponding time series data. At the same time, determine the normalized error bound of each channel through the upper bound of the maximum absolute error of each channel and the corresponding scaling factor.

[0010] S3: Calculate the correlation metric value between any two channels corresponding to the current sliding window based on the normalized channel time series data corresponding to the channel. When the correlation metric value is greater than or equal to the preset correlation threshold, establish an edge between the two channels and judge all channels one by one to construct a channel correlation graph.

[0011] S4: Determine the grouping threshold, group the channels in the channel correlation graph according to the grouping threshold, and select at least one channel in each channel group as the anchor channel, and the selected anchor channels constitute the anchor set.

[0012] S5: Construct a cross-channel correlation prediction model for non-anchor channels based on the anchor point set, and construct innovative sequences for non-anchor channels based on the cross-channel correlation prediction model;

[0013] S6: Determine the error sensitivity coefficient based on the cross-channel correlation prediction model, and then determine the quantification error bound of the anchor channel and the innovation sequence through the error sensitivity coefficient;

[0014] S7: Obtain the normalized reconstructed values ​​of the anchor point channel and the innovation sequence based on the quantization error bounds of the anchor point channel and the innovation sequence, respectively;

[0015] S8: Based on the normalized reconstruction value of the anchor channel and the normalized reconstruction value of the innovation sequence, the normalized reconstruction value of the non-anchor channel is obtained. Then, inverse normalization is performed according to the scale factor to obtain the reconstruction value of the channel under the original dimensions.

[0016] In a preferred embodiment of the present invention, in S2, statistical features are extracted from the time-series data of each channel within a set current sliding window to calculate the scaling factor, thereby eliminating the huge differences in dimensions and orders of magnitude between different types of physical quantities, as shown in the following formula.

[0017] ,

[0018] In the formula, As a scale factor, The size of the sliding window. Sampling time, It is a sliding window. This is channel timing data. For small bias terms;

[0019] The upper bound of the pointwise maximum absolute error adopts a mixed error bound, satisfying:

[0020] ,

[0021] In the formula, The minimum error value represents the minimum threshold for error. This is a relative error. This is channel timing data;

[0022] The time series data for each channel is normalized using the scaling factor and time series data for each channel. Simultaneously, the normalization error bound for each channel is determined using the upper bound of the maximum absolute error, as shown in the following formula.

[0023] ,

[0024] ,

[0025] In the formula, For any normalized channel time series data, This is channel timing data. As a scale factor, To normalize the error bound, This is the upper bound of the maximum absolute error point by point.

[0026] In a preferred embodiment of the present invention, in S3, the correlation metric is calculated using the absolute value of the Pearson correlation coefficient of the normalized time-series data of the two channels corresponding to the current sliding window, as shown in the following formula.

[0027] ,

[0028] In the formula, and These represent the channels corresponding to the current sliding window. and channels At any moment Normalized time series data, These represent the channels corresponding to the current sliding window. and channels The normalized time series mean, The size of the sliding window. To indicate the current sampling time, This is a measure of correlation.

[0029] The preset correlation threshold is determined based on the quantiles of the correlation set, as shown in the following formula.

[0030] ,

[0031] in The quantiles of the correlation values, This is a correlation metric. This is the correlation threshold;

[0032] when At that time, in the passage With channel Edges are established between channels, and each channel is evaluated one by one to construct a channel correlation graph.

[0033] In a preferred embodiment of the present invention, in S4, a maximum spanning tree or a maximum spanning forest is constructed in the channel correlation graph, and the grouping threshold is determined according to the quantile of the edge weight set in the maximum spanning tree or the maximum spanning forest, as shown in the following formula:

[0034] ,

[0035] In the formula, Indicates the grouping threshold; Represents the quantile function; denoted by quantile parameter; T represents the maximum spanning tree or maximum spanning forest constructed based on the channel correlation graph; e represents the edge in the maximum spanning tree or maximum spanning forest. Let e ​​represent the edge weight, where the edge weight is a correlation measure between the two corresponding channels;

[0036] When the edge weight of a side in the maximum spanning tree or maximum spanning forest is less than the grouping threshold, the edge is deleted. The channel correlation graph is divided into several unconnected subgraphs. Each subgraph is called a connected component, and each connected component is a channel group.

[0037] The bitrate estimation function is defined as follows:

[0038] ,

[0039] In the formula, This indicates that the normalized time series data sequence z is within the normalization error bound. The expected coding cost; This represents the standard deviation of the normalized time series data sequence z within the current sliding window W; W represents the current sliding window. This represents the normalized error bound corresponding to the normalized time series data sequence z; and This is a preset bitrate estimation constant;

[0040] Select a non-empty subset from the channel group as the anchor channel, and temporarily use it as the candidate anchor set. Construct a regression vector based on the normalized time series data of each channel in the candidate anchor set:

[0041] ,

[0042] In the formula, For the set of candidate anchor points The passage in the middle, This indicates the corresponding candidate anchor point channel at time [time]. Normalized time series data, constant term Used to introduce bias terms;

[0043] Within the current sliding window, the temporary prediction parameters of the non-anchor channels relative to the candidate anchor set are obtained using the least squares method, and the temporary prediction values ​​are obtained:

[0044] ,

[0045] In the formula, These are temporary prediction parameters;

[0046] The expected residual sequence of the non-anchor channel under the candidate anchor set is:

[0047] ,

[0048] In the formula, Indicates non-candidate anchor channel In the candidate anchor set The expected residual sequence below, Non-anchor channel At any moment Normalized time series data;

[0049] Candidate anchor set The corresponding channel group's estimated encoding cost is shown in the following formula:

[0050] ,

[0051] In the formula, Represents the set of candidate anchor points The corresponding channel group is expected to have a high encoding cost; Indicates candidate anchor channel Normalized time series data; Indicates candidate anchor channel The normalized error bound; Indicates non-anchor channel In the candidate anchor set The expected residual sequence obtained by temporary cross-channel prediction; Indicates non-anchor channel The normalized error bound;

[0052] Repeat the above operation, traversing all possible non-empty candidate anchor sets. The candidate anchor set that minimizes the expected encoding cost of the channel group is determined as the anchor set of the current channel group. The remaining channels within the corresponding channel group are the non-anchor channels, as shown in the following formula:

[0053] ,

[0054] In the formula, Indicates the current channel group The set of anchor points; Indicates from channel group The set of non-empty candidate channels selected from the data; Make the objective function The set of candidate channels corresponding to the minimum value.

[0055] In a preferred embodiment of the present invention, in S5, for non-anchor point channels, the construction time... The regression vector is as follows:

[0056] ,

[0057] In the formula, Indicates anchor point channel At any moment Normalized time series data, The constant term 1 is used to introduce a constant bias term. This represents a regression vector consisting of anchor channel normalized time series data and a constant term;

[0058] Non-anchor channel at time The model parameter vector is represented as:

[0059] = ,

[0060] In the formula, Indicates time Time Anchor Point Channel For non-anchor channels The regression coefficients; Indicates time Non-anchor channel The corresponding constant bias term;

[0061] Based on the aforementioned regression vector and model parameter vector, a cross-channel correlation prediction model for non-anchor channels is constructed. The non-anchor channel at time... The predicted value is:

[0062] ,

[0063] The innovation sequence of non-anchor channels is constructed based on the cross-channel correlation prediction model, as shown in the following formula:

[0064] ,

[0065] In the formula, Non-anchor channel At any moment The innovative sequence value represents the difference between the actual normalized time series data of the non-anchor channel and the cross-channel predicted value; Non-anchor channel At any moment The actual normalized time series data; Non-anchor channel At any moment The predicted value.

[0066] In a preferred embodiment of the present invention, in S6, the upper bound of the error sensitivity coefficient is defined as follows, based on the regression coefficients of the cross-channel prediction model:

[0067] ,

[0068] In the formula, Indicates anchor point channel The quantization error is transmitted to the non-anchor channel via the cross-channel prediction model. Maximum linear amplification factor at that time;

[0069] The quantization error bound of the anchor channel is defined as follows:

[0070] ,

[0071] In the formula, Indicates anchor point channel The quantization error bound used in subsequent quantization processes; For anchor point channels The normalized error bound; Non-anchor channel The normalized error bound; The number of anchor point channels in the anchor point set. This represents the anchor point error budget ratio.

[0072] For any non-anchor channel The upper bound for the quantization error propagation of the anchor set is:

[0073] ,

[0074] In the formula, The quantization error of the anchor set is transmitted to the non-anchor channel via the cross-channel prediction model. The upper bound of the maximum error is formed subsequently;

[0075] The remaining error after deducting the upper bound of the anchor point error propagation is used as the quantization error bound of the innovation sequence:

[0076] ,

[0077] In the formula, Indicates non-anchor channel Innovation sequence The quantization error bound used in subsequent quantization processes Non-anchor channel The normalized error bound.

[0078] In a preferred embodiment of the present invention, in S7, the historical normalized reconstructed value that can be used for the sequence to be quantized is composed of the results of quantization and reconstruction completed in the previous time step, expressed as:

[0079] ,

[0080] In the formula, The sequence to be quantized is either the normalized time series data of the anchor channel or the innovative sequence of the non-anchor channel. This indicates the sequence to be quantized at the previous time step. The normalized reconstructed values ​​that have already been obtained;

[0081] Based on the historical normalized reconstructed values, the causal prediction value of the sequence to be quantized at time t is calculated, as shown in the following formula:

[0082] ,

[0083] In the formula, For causal prediction values, For the first Prediction coefficients, , This indicates the sequence to be quantized at the previous time step. The normalized reconstructed values ​​that have already been obtained;

[0084] The prediction residual of the sequence to be quantized at time t is calculated as shown in the following formula:

[0085] ,

[0086] In the formula, To predict residuals.

[0087] When the sequence to be quantized For anchor point channels When normalizing time series data, the quantization error bound is:

[0088] ,

[0089] When the sequence to be quantized Non-anchor channel Innovation sequence When the quantization error bound is:

[0090] ,

[0091] In the formula, For anchor point channel quantization error bound, Innovative sequence quantization error bounds for non-anchor channels;

[0092] The quantization step size is set based on the quantization error bound. The quantization step size is selected from the power set and satisfies the following formula:

[0093] ,

[0094] In the formula, This is the quantization step size;

[0095] The predicted residuals are quantized into integers, as shown in the following formula:

[0096] ,

[0097] In the formula, For a moment The corresponding quantized integer, To predict residuals.

[0098] The normalized reconstructed value of the prediction residual is obtained from the quantized integer, as shown in the following formula:

[0099] ,

[0100] The normalized reconstructed value of the sequence to be quantized at time step [time] is then obtained, as shown in the following formula:

[0101] ,

[0102] In the formula, The sequence to be quantized At any moment The normalized reconstructed value, For causal prediction values, Normalized reconstructed value of the residual.

[0103] In a preferred embodiment of the present invention, in S8, the normalized time-series data of the anchor channel is replaced with the normalized reconstructed value of the anchor channel to obtain the reconstructed regression vector:

[0104] ,

[0105] In the formula, For the set of anchor points Each anchor point channel in the middle, For anchor point channels The normalized reconstructed value;

[0106] Based on the established cross-channel correlation prediction model, the normalized prediction part of the non-anchor channel is obtained by using the model parameter vector and the reconstructed regression vector. This normalized prediction part is then superimposed with the normalized reconstructed value of the innovation sequence to obtain the normalized reconstructed value of the non-anchor channel.

[0107] ,

[0108] In the formula, These are the normalized reconstructed values ​​for non-anchor channels. Non-anchor channel At any moment The model parameter vector, Construct a regression vector. Non-anchor channel The value of the innovation sequence reconstruction;

[0109] After obtaining the normalized reconstructed values ​​for each channel, inverse normalization is performed based on the scale factor to obtain the reconstructed values ​​in the original dimensions:

[0110] ,

[0111] In the formula, Indicates any channel. For channel The scale factor, For channel The normalized reconstructed value, For channel Reconstructed values ​​under the original dimensions.

[0112] The second objective of this invention is to provide a multi-channel time-series data error boundary quantization system based on dynamic graph grouping and error allocation, employing the multi-channel time-series data error boundary quantization method based on dynamic graph grouping and error allocation described above. The system includes a memory for storing a computer program and a processor for executing the computer program to implement the aforementioned multi-channel time-series data error boundary quantization method.

[0113] The processor includes a data preprocessing and normalization module, a graph correlation dynamic grouping module, a sensitivity coefficient and error budgeting module, and a quantization encoding and closed-loop reconstruction module;

[0114] The data preprocessing and normalization module is used to calculate the channel scale factor in real time and complete the dimensional unification and error boundary mapping of heterogeneous multi-channel data;

[0115] The graph correlation dynamic grouping module is used to update the spatial topology graph between multiple channels in real time with a sliding window and to perform minimum bit rate driven anchor point selection.

[0116] The sensitivity coefficient and error budget module is used to update the parameters of the cross-channel prediction model online, extract innovative sequences, and perform dynamic error budget splitting between anchor points and non-anchor points based on strict inequality mathematical constraints.

[0117] The quantization encoding and closed-loop reconstruction module is used to generate and encode an error-controlled integer stream, while the built-in event-triggered state machine enables robust switching between multiple modes.

[0118] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0119] This invention breaks through the limitation of traditional error bound compression methods that isolate each sensor channel, and achieves the ultimate exploitation of multi-channel spatial redundancy and a significant improvement in the overall compression ratio.

[0120] By dynamically calculating multi-channel correlation maps within a sliding window, dividing the optimal anchor point set, and constructing a cross-channel prediction model, this invention transforms the original high-entropy data into an innovative sequence with minimal variance, fundamentally eliminating mechanistic redundancy and information overlap between different physical quantities at the same collection point.

[0121] Meanwhile, addressing the technical pain point of "error propagation and step-by-step amplification" that is prone to occur in cross-channel prediction, this invention breaks away from the blindness of traditional static error allocation. It is the first to derive and introduce the "upper bound of the sensitivity coefficient" of the prediction model to the anchor point error, achieving a true absolute guarantee of deterministic error. Through strict mathematical inequality constraints, the globally allowed error limit is scientifically and rigorously allocated to the innovative sequence of anchor point channels and non-anchor point channels, ensuring that even if the reference anchor point itself has quantization error, the cumulative error after the reconstruction of the non-anchor point channel will never exceed the strict bottom line set by the user.

[0122] This enables extreme compression and high restoration of the original data, taking into account both the real-time processing needs of low-computing-power edge devices and strict error protection requirements. It can provide reliable technical support for lossy compression of multi-channel time-series data in industrial IoT scenarios, and can still stably achieve high compression ratio and strictly controllable error data compression processing on edge terminals with limited bandwidth and storage resources. Attached Figure Description

[0123] Figure 1 This is a flowchart of the multi-channel time-series data error boundary quantization method based on dynamic graph grouping and error allocation according to the present invention. Detailed Implementation

[0124] Typical embodiments embodying the features and advantages of the present invention will be specifically described in the following description. It should be understood that the present invention can have various variations in different embodiments without departing from the scope of the present invention, and the descriptions and illustrations therein are for illustrative purposes only and not intended to limit the present invention.

[0125] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0126] This embodiment discloses a method for error bound quantization of multi-channel time-series data based on dynamic graph grouping and error allocation, including the following steps:

[0127] S1: Acquire multi-channel time-series data collected from the same acquisition point. ,in For the number of channels, Sampling time;

[0128] In this embodiment, the multi-channel timing data is voltage. ,vibration Data collected by the corresponding sensors.

[0129] S2: For each channel Determine the scale factor And determine each channel Upper bound of pointwise maximum absolute error under original dimensions And by using the scaling factor and time series data of each channel, the time series data of each channel are normalized to Simultaneously, the normalized error bound of each channel is determined by the upper bound of the maximum absolute error of each channel. As shown in the formula below,

[0130] ,

[0131] ,

[0132] In the formula, For any normalized channel time series data, This is channel timing data. As a scale factor, To normalize the error bound, This is the upper bound of the maximum absolute error point by point.

[0133] In this embodiment, the timing data of each channel is set within a current sliding window. Statistical feature extraction is performed on the data to calculate the scaling factor, thereby eliminating the significant differences in dimensions and orders of magnitude between different types of physical quantities, as shown in the following formula.

[0134] ,

[0135] In the formula, As a scale factor, The size of the sliding window. Sampling time, It is a sliding window. This is channel timing data. To introduce a small bias term, a small bias term is introduced. To prevent division by zero errors caused by dead zone data.

[0136] The upper bound of the pointwise maximum absolute error adopts a mixed error bound, which satisfies This allows large signals to tolerate larger absolute errors, while small signals maintain a strict baseline. The minimum error value represents the minimum threshold for error. This is a relative error. This is channel timing data.

[0137] When the signal amplitude is small, it is determined by the absolute bottom line. To prevent noise from drowning out the signal; relative error is permissible when the signal experiences pulses or large fluctuations. Interventions were made to improve compression efficiency, ultimately resulting in a unified dimensionless normalized error bound. .

[0138] This embodiment completes data acquisition and preprocessing through S1 and S2, eliminating the huge differences in dimensions and orders of magnitude between data collected by different sensors in the industrial field, and providing a unified scalar field for subsequent graph calculation and error allocation.

[0139] S3: For any two channels corresponding to the current sliding window and The correlation metric between any two channels is calculated based on their normalized channel time series data. When the correlation metric is greater than or equal to a preset correlation threshold, an edge is established between the two channels to construct a channel correlation graph. .

[0140] In this embodiment, Indicates channel and channels Correlation measure between them Use the channel corresponding to the current sliding window With channel The absolute value of the Pearson correlation coefficient of the normalized time series data is calculated as shown in the following formula.

[0141] ,

[0142] In the formula, and These represent the channels corresponding to the current sliding window. and channels At any moment Normalized time series data, These represent the channels corresponding to the current sliding window. and channels The normalized time series mean, The size of the sliding window. This represents the current sampling time, which is the end time of the current sliding window.

[0143] The range of values ​​is The larger the value, the more channel With channel The stronger the correlation.

[0144] Specifically, the preset correlation threshold will be used. Determined by the quantiles of the correlation set, as shown in the following formula.

[0145] ,

[0146] in The quantiles of the correlation values, This is a correlation metric used to adaptively filter out weakly correlated noise edges.

[0147] In this embodiment, a quantile threshold is used to avoid computational explosion caused by fully connected graphs. Dynamically truncate weakly correlated edges to improve the channel correlation graph. It can adaptively filter out spurious correlations caused by environmental noise; specifically, when At that time, in the passage With channel Establish edges between channels, and evaluate each channel one by one (traversing all channels) to construct a channel correlation graph.

[0148] In this embodiment, the node set The set of all channels, edge set express The set of edges.

[0149] In this embodiment, let there be a total Each channel, iterating through all channel pairs. ,in For each pair of channels and Calculate its correlation measure. When At that time, in the passage and Establish an edge between each channel. After traversing all channels, construct a channel correlation graph using all channels as nodes and all established edges as edge sets. .

[0150] When the channel correlation metric value within the current sliding window is lower than the preset correlation threshold, or when an innovative sequence... When the fluctuation exceeds the preset threshold, it is determined that the prediction relationship between the corresponding non-anchor channel and the current anchor set is weakened, and the channel is switched to the independent error bound quantization processing mode. When the channel correlation metric recovers and meets the preset requirements, the dynamic graph grouping, anchor selection and cross-channel prediction process are re-executed to ensure the reconstruction accuracy and robustness under dynamic conditions.

[0151] Through the above switching mechanism, the system can promptly revert to the single-channel error bound quantization mode when cross-channel prediction fails or correlation changes abruptly, thereby ensuring that the entire system still has good robustness and error controllability under dynamic operating conditions.

[0152] This invention also overcomes the vulnerability of static compression strategies to non-stationary industrial data, achieving strong robustness and hardware-friendly adaptive state switching. By combining data normalization processing with an event triggering mechanism based on data change rate, when a step change in equipment operating conditions leads to a drastic change in the correlation graph, or when cross-channel prediction gains decrease, the system can quickly and seamlessly revert to independent single-channel error bound quantization processing, thereby perfectly ensuring the algorithm's extreme stability in non-stationary industrial time-series data streams and high availability on resource-constrained edge devices.

[0153] S4: Determine the grouping threshold, group the channels in the channel correlation graph according to the grouping threshold, and select at least one channel in each channel group as the anchor channel, and the selected anchor channels constitute the anchor set.

[0154] In this embodiment, when the channel correlation graph is connected (in this graph composed of channels as nodes and correlations as edges, any channel can reach any other channel directly or indirectly along the correlation edge), a maximum spanning tree is constructed in the channel correlation graph. Specifically, the correlation metric values ​​between all channels are sorted in descending order, and the pair of channels corresponding to the maximum current correlation metric value are selected as the edge to be connected. If the edge does not form a cycle with the selected edge, it is included in the maximum spanning tree.

[0155] When the channel correlation graph itself is not connected, the channel correlation graph includes multiple connected subgraphs. Then, a maximum spanning tree is constructed for each connected subgraph, and multiple maximum spanning trees form a maximum spanning forest. Specifically, the above maximum spanning tree construction process is performed on each connected subgraph, and finally the maximum spanning forest corresponding to the entire channel correlation graph is obtained.

[0156] The grouping threshold is determined based on the set of edge weights in the maximum spanning tree or the maximum spanning forest, and edges with edge weights less than the grouping threshold are deleted, thereby dividing the channel correlation graph into several channel groups.

[0157] Specifically, the grouping threshold is determined according to the quantiles of the edge weight set in the maximum spanning tree or the maximum spanning forest, as shown in the following formula:

[0158] ,

[0159] In the formula, Indicates the grouping threshold; Represents the quantile function; denoted by quantile parameter; T represents the maximum spanning tree or maximum spanning forest constructed based on the channel correlation graph; e represents the edge in the maximum spanning tree or maximum spanning forest. Let e ​​represent the edge weight, where the edge weight is a correlation metric between the two corresponding channels.

[0160] When the edge weight of a certain side in a maximum spanning tree or maximum spanning forest Less than the grouping threshold When the edge is deleted, the channel correlation graph is divided into several unconnected subgraphs. Each subgraph is called a connected component, and each connected component is a channel group.

[0161] In this embodiment, edges with weights less than the grouping threshold are deleted, which means that the edges between any two channels with correlation metrics less than the grouping threshold are deleted.

[0162] Therefore, channels within the same channel group have a strong correlation, while the correlation between different channel groups is relatively weak.

[0163] In the above manner, the correlation threshold in S3 is used to filter out globally weakly correlated edges and construct a channel correlation graph, and the grouping threshold in S4 is used to further prune on the maximum spanning tree or the maximum spanning forest to obtain a more compact channel grouping structure.

[0164] For any channel group, at least one channel in the channel group is selected as an anchor point channel, and the selected anchor point channels constitute the anchor point set. .

[0165] The anchor channel is used as a reference channel for cross-channel prediction in subsequent steps, so it should be able to represent the main changing trends within the channel group and help reduce the prediction coding cost of the entire channel group.

[0166] This embodiment uses the minimum predicted bitrate criterion to determine the anchor point set. First, the bitrate estimation function is defined as shown in the following formula:

[0167] ,

[0168] In the formula, This indicates that the normalized time series data sequence z is within the normalization error bound. The expected coding cost; This represents the standard deviation of the normalized time series data sequence z within the current sliding window W; W represents the current sliding window. This represents the normalized error bound corresponding to the normalized time series data sequence z; and This is a preset bitrate estimation constant.

[0169] Generally speaking, the greater the sequence fluctuation, the larger its standard deviation, and the higher the expected coding cost; the larger the allowable error bound, the lower the expected coding cost.

[0170] From the channel group Select a non-empty subset from As an anchor point channel, and temporarily as a set of candidate anchor points, and If a channel belongs to the set of candidate anchor points ,Right now If a channel belongs to a channel group, then that channel is directly encoded as a candidate anchor channel; However, it does not belong to the candidate anchor point set. ,Right now Then use the candidate anchor set The normalized time series data of each channel in the middle are used to perform temporary cross-channel prediction for this channel. Specifically, let the current channel group be... The candidate anchor set is ,and , For any non-candidate anchor point channel Based on the candidate anchor set The regression vector is constructed from the normalized time series data of each channel:

[0171] ,

[0172] In the formula, For the set of candidate anchor points The passage in the middle, This indicates the corresponding candidate anchor point channel at time [time]. Normalized time series data, constant term Used to introduce bias terms.

[0173] In the current sliding window Within, the non-anchor point channel is obtained using the least squares method. Relative to the set of candidate anchor points Temporary prediction parameters And obtain the provisional predicted value:

[0174] ,

[0175] Non-anchor channel In the candidate anchor set The expected residual sequence is as follows:

[0176] ,

[0177] In the formula, Indicates non-candidate anchor channel In the candidate anchor set The expected residual sequence below, Non-anchor channel At any moment Normalized time series data;

[0178] Candidate anchor set The corresponding channel group's estimated encoding cost is shown in the following formula:

[0179] ,

[0180] In the formula, Represents the set of candidate anchor points The corresponding channel group is expected to have a high encoding cost; Indicates candidate anchor channel Normalized time series data; Indicates candidate anchor channel The normalized error bound; Indicates non-anchor channel In the candidate anchor set The expected residual sequence obtained by temporary cross-channel prediction; Indicates non-anchor channel The normalized error bound.

[0181] In the above formula, the first term represents the expected encoding cost required for direct encoding of the candidate anchor channel itself; the second term represents the expected encoding cost required for encoding the expected residual sequence after the other channels in the channel group perform cross-channel prediction using the candidate anchor set. By considering both costs simultaneously, the problem of selecting anchors solely based on correlation strength while ignoring the overall compression effect can be avoided.

[0182] Repeat the above operation to traverse all possible non-empty candidate anchor sets. The candidate anchor set that minimizes the expected encoding cost of the channel group will be determined as the anchor set of the current channel group, and the remaining channels in the corresponding channel group will be the non-anchor channels.

[0183] The set of candidate anchors that minimizes the expected encoding cost of the channel group is determined as the anchor set of the current channel group, as shown in the following formula:

[0184] ,

[0185] In the formula, Indicates the current channel group The set of anchor points; Indicates from channel group The set of non-empty candidate channels selected from the data; Make the objective function The set of candidate channels corresponding to the minimum value.

[0186] In this approach, instead of pre-specifying fixed anchor points, this step adaptively selects at least one channel as the anchor channel within each channel group based on the channel correlation structure and expected encoding cost within the current sliding window. The determined set of anchor points... This will serve as the input for establishing a cross-channel correlation prediction model in S5; for the remaining channels within the channel group that are not selected as anchor channels, the corresponding innovation sequence will be constructed in S5 based on the anchor set.

[0187] S5: Construct a cross-channel correlation prediction model for non-anchor channels based on the anchor set, and construct innovative sequences for non-anchor channels based on the cross-channel correlation prediction model.

[0188] For each channel group Any non-anchor channel in Based on anchor point set A cross-channel correlation prediction model is established using normalized time series data to the non-anchor channel, in order to eliminate data redundancy by utilizing the correlation between channels within the channel group.

[0189] Specifically, let the set of anchor points be: ,

[0190] Where m is the number of anchor point channels. These represent the channels at each anchor point.

[0191] For non-anchor channel Construction time The regression vector is as follows:

[0192] ,

[0193] In the formula, Indicates anchor point channel At any moment Normalized time series data, The constant term 1 is used to introduce a constant bias term. This represents the regression vector consisting of anchor channel normalized time series data and a constant term.

[0194] Non-anchor channel At any moment The model parameter vector is represented as:

[0195] = ,

[0196] In the formula, Indicates time Time Anchor Point Channel For non-anchor channels The regression coefficients; Indicates time Non-anchor channel The corresponding constant bias term.

[0197] Based on the aforementioned regression vector and model parameter vector, a cross-channel correlation prediction model for non-anchor channels is constructed. The predicted value at time t is:

[0198] ,

[0199] When expanded, it can be represented as:

[0200] ,

[0201] As can be seen from the above formula, the model parameter vector It also includes the regression coefficients of each anchor point channel to the non-anchor point channel. and constant bias term .

[0202] Constructing non-anchor channels based on cross-channel correlation prediction model The innovation sequence is shown in the following formula:

[0203] ,

[0204] In the formula, Non-anchor channel At any moment The innovative sequence value represents the difference between the actual normalized time series data of the non-anchor channel and the cross-channel predicted value; Non-anchor channel At any moment The actual normalized time series data; Non-anchor channel At any moment The predicted value.

[0205] Innovation sequences characterize the set of anchor points Beyond the explained relevant components, this non-anchor channel contains remaining unexplained information.

[0206] The synchronous update of the model parameter vector is performed in S8 based on the reconstructed values ​​at the current time step to ensure that the encoder and decoder use consistent model parameters in subsequent time steps.

[0207] S6: Determine the quantization error bounds of the anchor channel and the innovation sequence based on the error sensitivity coefficient.

[0208] In S5, it is already based on the anchor point set. A cross-channel correlation prediction model for non-anchor channels is established, and the non-anchor channels are obtained. Innovation sequence Since anchor channels generate reconstruction errors after quantization, these errors are propagated to non-anchor channels through the cross-channel prediction model. Therefore, it is necessary to determine the error bounds used for anchor channels and innovative sequences in subsequent quantization processes based on the error sensitivity of the prediction model.

[0209] Let the set of non-anchor channels in the current channel group be:

[0210] ,

[0211] For any non-anchor channel and any anchor point channel Based on the regression coefficients of the cross-channel prediction model in S5 The upper bound of the error sensitivity coefficient is defined as:

[0212] ,

[0213] In the formula, Indicates anchor point channel The quantization error is transmitted to the non-anchor channel via the cross-channel prediction model. The maximum linear amplification factor at that time.

[0214] Let the anchor point error budget ratio be . ,and To ensure that the quantization error of the anchor channel, after propagation through the prediction model, does not exceed a portion of the normalization error bound of the non-anchor channel, the anchor channel... The quantization error bound is defined as follows:

[0215] ,

[0216] In the formula, Indicates anchor point channel The quantization error bound used in subsequent quantization processes; For anchor point channels The normalized error bound obtained in S2; Non-anchor channel The normalized error bound obtained in S2; This represents the number of anchor point channels in the anchor point set.

[0217] when When, it indicates the anchor point channel. The error will not be propagated to the non-anchor channel through this prediction model. This item is not included in the above minimum value calculation.

[0218] For any non-anchor channel The upper bound for the quantization error propagation of the anchor set is:

[0219] ,

[0220] In the formula, The quantization error of the anchor set is transmitted to the non-anchor channel via the cross-channel prediction model. The upper bound of the maximum error is then formed.

[0221] To ensure non-anchor point channels The final reconstruction error does not exceed its normalized error bound. The remaining error after deducting the upper bound of the anchor point error propagation is taken as the innovation sequence. Quantization error bound:

[0222] ,

[0223] In the formula, Indicates non-anchor channel Innovation sequence The quantization error bound used in the subsequent quantization process.

[0224] when If this indicates that the anchor point error propagation has consumed too much of the error budget, then the proportion of the anchor point error budget should be reduced. Alternatively, the corresponding non-anchor channel can be switched to independent error bound quantization processing mode.

[0225] Using the above method, this step outputs the anchor point channel quantization error boundary. And innovative sequence quantization error boundary .in, This is used for subsequent quantization of the anchor point channel normalized time series data. Used for subsequent innovation sequences of non-anchor point channels Quantify it.

[0226] S7: Perform integer quantization of the prediction residuals guaranteed by the execution error bound, and obtain the normalized reconstructed values ​​of the anchor point channel and the innovation sequence according to the quantization error bounds of the anchor point channel and the innovation sequence, respectively.

[0227] Causal prediction, residual quantization, and loop closure reconstruction are performed on the sequence to be quantized. The sequence to be quantized is either normalized time-series data from an anchor channel or an innovative sequence from a non-anchor channel. For ease of description, the sequence to be quantized is denoted as... .

[0228] Specifically, let the prediction order be... For the current moment Quantization sequence The usable historical normalized reconstructed values ​​consist of the results of quantization and reconstruction completed in previous time steps, expressed as:

[0229] ,

[0230] Among them, any historical normalized reconstructed value All are derived from the corresponding preceding time. The predicted value and the normalized reconstructed value of the quantized residual are obtained as shown in the following formula:

[0231] ,

[0232] ,

[0233] ,

[0234] In the formula, , This indicates the sequence to be quantized at the previous time step. The normalized reconstructed values ​​that have already been obtained, Indicates the preceding time. causal predictive value, Indicates the preceding time. Quantized integers, This indicates the quantization step size corresponding to the sequence to be quantized.

[0235] Therefore, at the current moment The historical normalized reconstruction values ​​used are all derived from the closed-loop quantization reconstruction results of the preceding time step and are synchronously saved at the encoding and decoding ends.

[0236] Based on the historical normalized reconstructed values, the sequence to be quantized is calculated. At any moment The causal predictive value is shown in the following formula:

[0237] ,

[0238] In the formula, For causal prediction values, For the first Prediction coefficients.

[0239] Calculate the sequence to be quantized At any moment The predicted residuals are shown in the following formula:

[0240] ,

[0241] In the formula, To predict residuals.

[0242] Suppose the sequence to be quantized The corresponding quantization error bound is When the sequence to be quantized For anchor point channels When normalizing time series data, take:

[0243] ,

[0244] When the sequence to be quantized Non-anchor channel Innovation sequence At that time, take:

[0245] ,

[0246] In the formula, The anchor point channel quantization error bound is determined for S6. The quantization error bound for the non-anchor channel innovation sequence is determined for S6.

[0247] Based on the quantization error bound Set quantization step size The quantization step size is selected from the exponentiation set and satisfies the following formula:

[0248] ,

[0249] Predicting residuals Quantized to integers, as shown in the following formula:

[0250] ,

[0251] In the formula, For a moment The corresponding quantized integer.

[0252] Based on quantized integers The normalized reconstructed value of the predicted residual is obtained as shown in the following formula:

[0253] ,

[0254] This leads to the sequence to be quantized. At any moment The normalized reconstructed value is shown in the following formula:

[0255] ,

[0256] As can be seen from the above quantization and reconstruction process, the sequence to be quantized At any moment The reconstruction error satisfies:

[0257] ,

[0258] This ensures that the normalized time series data of the anchor channel and the innovative sequence of the non-anchor channel are quantized and reconstructed under the quantization error bounds determined by S6.

[0259] Quantized integers obtained at each time point Perform entropy encoding or binary encoding to form a quantized integer sequence and output it; simultaneously, reconstruct the current normalized value. As the historical normalized reconstruction value for causal prediction at the next moment, it forms a closed-loop prediction quantization process with consistent encoding and decoding.

[0260] In the above manner, S7 uses the anchor channel quantization error boundary output by S6. And innovative sequence quantization error boundary As input, normalized time series data from anchor channels and innovative sequences from non-anchor channels were processed respectively. The prediction residual is quantized into integers, and the corresponding quantized integer sequence and normalized reconstructed value are output, which provides the basis for restoring the non-anchor channel in S8 based on the anchor point normalized reconstructed value and the innovation sequence normalized reconstructed value.

[0261] S8: Based on the normalized reconstruction value of the anchor channel and the normalized reconstruction value of the innovation sequence, the normalized reconstruction value of the non-anchor channel is obtained. Then, inverse normalization is performed according to the scale factor to obtain the reconstruction value of the channel under the original dimensions.

[0262] In S7, the normalized time series data of the anchor channel and the innovative sequences of the non-anchor channel have been quantized and reconstructed, respectively.

[0263] When the sequence to be quantized For anchor point channels Normalized time series data At that time, its normalized reconstructed value Recorded as: When the sequence to be quantized Non-anchor channel Innovation sequence At that time, its normalized reconstructed value Recorded as: .

[0264] For any non-anchor channel Following the construction method of the regression vector in S5.

[0265] Replace the normalized time series data of the anchor channel with the normalized reconstructed values ​​of the anchor channel to obtain the reconstructed regression vector:

[0266] ,

[0267] In the formula, For the set of anchor points Each anchor point channel in the middle, For anchor point channels Normalized reconstructed value; constant term It is consistent with the regression vector in S5.

[0268] Based on the cross-channel correlation prediction model established in S5, the model parameter vector is used. and reconstructing regression vectors Obtain the non-anchor channel The normalized prediction part is then superimposed with the reconstructed value of the innovation sequence to obtain the non-anchor channel. Normalized reconstructed values:

[0269] ,

[0270] In the formula, These are the normalized reconstructed values ​​for non-anchor channels. Non-anchor channel At any moment The model parameter vector, Construct a regression vector. Non-anchor channel The value of the innovation sequence reconstruction;

[0271] Therefore, the normalized reconstruction value of the non-anchor channel is obtained by jointly recovering the predicted part formed by the anchor channel reconstruction value and the innovation sequence reconstruction value.

[0272] Since the upper bound of anchor point error propagation has been determined in S6... And innovative sequence quantization error boundary And satisfy:

[0273] ,

[0274] Therefore, non-anchor channel The normalized reconstruction error does not exceed its normalized error bound. .

[0275] After obtaining the normalized reconstructed values ​​for each channel, inverse normalization is performed based on the scale factor determined in S2 to obtain the reconstructed values ​​under the original dimensions:

[0276] ,

[0277] In the formula, Indicates any channel. For channel The scale factor, For channel The normalized reconstructed value, For channel The reconstructed values ​​under the original dimensions. Therefore, the reconstruction error of each channel under the original dimensions does not exceed the upper bound of the corresponding point-by-point maximum absolute error.

[0278] In this embodiment, in order to maintain a consistent prediction model state between the encoder and decoder in subsequent time steps, both the encoder and decoder use the anchor channel reconstruction value that can be jointly obtained at the current time. Non-anchor channel reconstruction value The model parameter vector in S5 is updated according to the same rules as follows: Since the update process relies solely on the reconstructed values ​​available at both ends, it ensures that the encoding and decoding ends maintain a consistent cross-channel prediction model in the next time step.

[0279] In this embodiment, multi-channel time-series data is collected at the edge node (i.e., the encoding end), then processed to obtain innovative sequences of anchor channels and non-anchor channels, which is to obtain the corresponding compressed data. The compressed data is then uploaded to the cloud (decoding end), and the reconstructed values ​​of the channels under the original dimensions are obtained by decoding through quantization error bound and normalized reconstruction values.

[0280] As shown in Table 1 below, to verify the superiority of this invention, key technical indicators of traditional methods and this invention were compared on a multi-channel industrial time-series dataset containing strong non-stationarity and dynamic start-stop fluctuations. The experimental data conclusively demonstrate that this invention achieves a comprehensive generational performance advantage while maintaining high security and low overhead.

[0281] Table 1: Overall Performance Comparison of Multi-channel Time-series Data Quantization Compression Schemes

[0282] Quantization and Compression Methods Compression ratio (CR) Maximum point-by-point error guarantee rate Signal-to-noise ratio (SNR) under extreme operating conditions Memory overhead of a single node at the edge Traditional single-channel independent error bound compression 14.5 98.6% 31.2dB 8.5KB Static grouping cross-channel cross-prediction compression 26.8 83.4% 22.4dB (error amplification) 12.0KB Method of the present invention 35.2 98.4% 38.5dB 9.2KB

[0283] This embodiment also discloses a multi-channel time-series data error boundary quantization system based on dynamic graph grouping and error allocation, including a memory configured to store a computer program; and a processor configured to execute the computer program to implement the above-described multi-channel time-series data error boundary quantization method.

[0284] The processor includes a data preprocessing and normalization module, a graph correlation dynamic grouping module, a sensitivity coefficient and error budgeting module, and a quantization encoding and closed-loop reconstruction module.

[0285] The data preprocessing and normalization module is responsible for calculating the channel scale factor in real time and completing the dimensional unification and error boundary mapping of heterogeneous multi-channel data.

[0286] The graph correlation dynamic grouping module has a built-in Pearson or cross-correlation calculation engine, which is responsible for updating the spatial topology graph between multiple channels in real time with a sliding window and performing minimum bit rate driven anchor point selection.

[0287] The sensitivity coefficient and error budget module is responsible for updating the parameters of the cross-channel prediction model online, extracting innovative sequences, and performing dynamic error budget segmentation between anchor points and non-anchor points based on strict inequality mathematical constraints.

[0288] The quantization encoding and closed-loop reconstruction module includes an adaptive causal predictor and an integer quantizer, which are responsible for generating and encoding an error-controlled integer stream. It also has a built-in event-triggered state machine to achieve robust switching between multiple modes.

[0289] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for error bound quantization of multi-channel time-series data based on dynamic graph grouping and error allocation, characterized in that, Includes the following steps: S1: Acquire multi-channel time-series data collected from the same acquisition point; S2: Determine the scaling factor for each channel and the upper bound of the maximum absolute error of each channel under the original dimensions. Obtain the normalized channel time series data through the scaling factor of each channel and the corresponding time series data. At the same time, determine the normalized error bound of each channel through the upper bound of the maximum absolute error of each channel and the corresponding scaling factor. S3: Calculate the correlation metric value between any two channels corresponding to the current sliding window based on the normalized channel time series data corresponding to the channel. When the correlation metric value is greater than or equal to the preset correlation threshold, establish an edge between the two channels and judge all channels one by one to construct a channel correlation graph. S4: Determine the grouping threshold, group the channels in the channel correlation graph according to the grouping threshold, and select at least one channel in each channel group as the anchor channel, and the selected anchor channels constitute the anchor set. S5: Construct a cross-channel correlation prediction model for non-anchor channels based on the anchor point set, and construct innovative sequences for non-anchor channels based on the cross-channel correlation prediction model; S6: Determine the error sensitivity coefficient based on the cross-channel correlation prediction model, and then determine the quantification error bound of the anchor channel and the innovation sequence through the error sensitivity coefficient; S7: Obtain the normalized reconstructed values ​​of the anchor point channel and the innovation sequence based on the quantization error bounds of the anchor point channel and the innovation sequence, respectively; S8: Based on the normalized reconstruction value of the anchor channel and the normalized reconstruction value of the innovation sequence, the normalized reconstruction value of the non-anchor channel is obtained. Then, inverse normalization is performed according to the scale factor to obtain the reconstruction value of the channel under the original dimensions.

2. The multi-channel time-series data error boundary quantization method based on dynamic graph grouping and error allocation according to claim 1, characterized in that: In S2, statistical features are extracted from the time-series data of each channel within the set current sliding window to calculate the scaling factor, thereby eliminating the huge differences in dimensions and orders of magnitude between different types of physical quantities, as shown in the following formula. ， In the formula, As a scale factor, The size of the sliding window. Sampling time, It is a sliding window. This is channel timing data. For small bias terms; The upper bound of the pointwise maximum absolute error adopts a mixed error bound, satisfying: ， In the formula, The minimum error value represents the minimum threshold for error. This is a relative error. This is channel timing data; The time series data for each channel is normalized using the scaling factor and time series data for each channel. Simultaneously, the normalization error bound for each channel is determined using the upper bound of the maximum absolute error, as shown in the following formula. ， ， In the formula, For any normalized channel time series data, This is channel timing data. As a scale factor, To normalize the error bound, This is the upper bound of the maximum absolute error point by point.

3. The multi-channel time-series data error boundary quantization method based on dynamic graph grouping and error allocation according to claim 1, characterized in that: In S3, the correlation metric is calculated using the absolute value of the Pearson correlation coefficient of the normalized time-series data of the two channels corresponding to the current sliding window, as shown in the following formula. ， In the formula, and These represent the channels corresponding to the current sliding window. and channels At any moment Normalized time series data, These represent the channels corresponding to the current sliding window. and channels The normalized time series mean, The size of the sliding window. To indicate the current sampling time, This is a measure of correlation. The preset correlation threshold is determined based on the quantiles of the correlation set, as shown in the following formula. ， in The quantiles of the correlation values, This is a correlation metric. This is the correlation threshold; when At that time, in the passage With channel Edges are established between channels, and each channel is evaluated one by one to construct a channel correlation graph.

4. The multi-channel time-series data error boundary quantization method based on dynamic graph grouping and error allocation according to claim 1, characterized in that: In S4, a maximum spanning tree or a maximum spanning forest is constructed in the channel correlation graph. The grouping threshold is determined according to the quantile of the edge weight set in the maximum spanning tree or the maximum spanning forest, as shown in the following formula: ， In the formula, Indicates the grouping threshold; Represents the quantile function; denoted by quantile parameter; T represents the maximum spanning tree or maximum spanning forest constructed based on the channel correlation graph; e represents the edge in the maximum spanning tree or maximum spanning forest. Let e ​​represent the edge weight, where the edge weight is a correlation measure between the two corresponding channels; When the edge weight of one side in the maximum spanning tree or maximum spanning forest is less than the grouping threshold, the edge is deleted. The channel correlation graph is divided into several unconnected subgraphs. Each subgraph is called a connected component, and each connected component is a channel group. The bitrate estimation function is defined as follows: ， In the formula, This indicates that the normalized time series data sequence z is within the normalization error bound. The expected coding cost; This represents the standard deviation of the normalized time series data sequence z within the current sliding window W; W represents the current sliding window; This represents the normalized error bound corresponding to the normalized time series data sequence z; and This is a preset bitrate estimation constant; Select a non-empty subset from the channel group as the anchor channel, and temporarily use it as the candidate anchor set. Construct a regression vector based on the normalized time series data of each channel in the candidate anchor set: ， In the formula, For the set of candidate anchor points The passage in the middle, This indicates the corresponding candidate anchor point channel at time [time]. Normalized time series data, constant term Used to introduce bias terms; Within the current sliding window, the temporary prediction parameters of the non-anchor channels relative to the candidate anchor set are obtained using the least squares method, and the temporary prediction values ​​are obtained: ， In the formula, These are temporary prediction parameters; The expected residual sequence of the non-anchor channel under the candidate anchor set is: ， In the formula, Indicates non-candidate anchor channel In the candidate anchor set The expected residual sequence below, Non-anchor channel At any moment Normalized time series data; Candidate anchor set The corresponding channel group's expected encoding cost is shown in the following formula: ， In the formula, Represents the set of candidate anchor points The corresponding channel group is expected to have a high encoding cost; Indicates candidate anchor channel Normalized time series data; Indicates candidate anchor channel The normalized error bound; Indicates non-anchor channel In the candidate anchor set The expected residual sequence obtained by temporary cross-channel prediction; Indicates non-anchor channel The normalized error bound; Repeat the above operation, traversing all possible non-empty candidate anchor sets. The candidate anchor set that minimizes the expected encoding cost of the channel group is determined as the anchor set of the current channel group. The remaining channels within the corresponding channel group are the non-anchor channels, as shown in the following formula: , In the formula, Indicates the current channel group The set of anchor points; Indicates from channel group The set of non-empty candidate channels selected from the data; Make the objective function The set of candidate channels corresponding to the minimum value.

5. The multi-channel time-series data error boundary quantization method based on dynamic graph grouping and error allocation according to claim 1, characterized in that: In S5, for non-anchor channels, the construction time is... The regression vector is as follows: ， In the formula, Indicates anchor point channel At any moment Normalized time series data, The constant term 1 is used to introduce a constant bias term. This represents a regression vector consisting of anchor channel normalized time series data and a constant term; Non-anchor channel at time The model parameter vector is represented as: = , In the formula, Indicates time Time Anchor Point Channel For non-anchor channels The regression coefficients; Indicates time Non-anchor channel The corresponding constant bias term; Based on the aforementioned regression vector and model parameter vector, a cross-channel correlation prediction model for non-anchor channels is constructed. The non-anchor channel at time... The predicted value is: , The innovation sequence of non-anchor channels is constructed based on the cross-channel correlation prediction model, as shown in the following formula: ， In the formula, Non-anchor channel At any moment The innovative sequence value represents the difference between the actual normalized time series data of the non-anchor channel and the cross-channel predicted value; Non-anchor channel At any moment The actual normalized time series data; Non-anchor channel At any moment The predicted value.

6. The multi-channel time-series data error boundary quantization method based on dynamic graph grouping and error allocation according to claim 1, characterized in that: In S6, based on the regression coefficients of the cross-channel prediction model, the upper bound of the error sensitivity coefficient is defined as follows: ， In the formula, Indicates anchor point channel The quantization error is transmitted to the non-anchor channel via the cross-channel prediction model. Maximum linear amplification factor at that time; The quantization error bound of the anchor channel is defined as follows: ， In the formula, Indicates anchor point channel The quantization error bound used in subsequent quantization processes; For anchor point channels The normalized error bound; Non-anchor channel The normalized error bound; The number of anchor point channels in the anchor point set. This represents the anchor point error budget ratio. For any non-anchor channel The upper bound for the quantization error propagation of the anchor set is: ， In the formula, The quantization error of the anchor set is transmitted to the non-anchor channel via the cross-channel prediction model. The upper bound of the maximum error is formed subsequently; The remaining error after deducting the upper bound of the anchor point error propagation is used as the quantization error bound of the innovation sequence: ， In the formula, Indicates non-anchor channel Innovation sequence The quantization error bound used in subsequent quantization processes Non-anchor channel The normalized error bound.

7. The multi-channel time-series data error boundary quantization method based on dynamic graph grouping and error allocation according to claim 1, characterized in that: In S7, the historical normalized reconstructed values ​​that can be used for the sequence to be quantized are composed of the results of quantization and reconstruction completed in the previous time step, expressed as: ， In the formula, The sequence to be quantized is either the normalized time series data of the anchor channel or the innovative sequence of the non-anchor channel. This indicates the sequence to be quantized at the previous time step. The normalized reconstructed values ​​that have already been obtained; Based on the historical normalized reconstructed values, the causal prediction value of the sequence to be quantized at time t is calculated, as shown in the following formula: ， In the formula, For causal prediction values, For the first Prediction coefficients, , This indicates the sequence to be quantized at the previous time step. The normalized reconstructed values ​​that have already been obtained; The prediction residual of the sequence to be quantized at time t is calculated as shown in the following formula: ， In the formula, To predict residuals; When the sequence to be quantized For anchor point channels When normalizing time series data, the quantization error bound is: ， When the sequence to be quantized Non-anchor channel Innovation sequence When the quantization error bound is: ， In the formula, For anchor point channel quantization error bound, Innovative sequence quantization error bounds for non-anchor channels; The quantization step size is set based on the quantization error bound. The quantization step size is selected from the power set and satisfies the following formula: ， In the formula, This is the quantization step size; The predicted residuals are quantized into integers, as shown in the following formula: ， In the formula, For a moment The corresponding quantized integer, To predict residuals; The normalized reconstructed value of the prediction residual is obtained from the quantized integer, as shown in the following formula: ， The normalized reconstructed value of the sequence to be quantized at time step [time] is then obtained, as shown in the following formula: ， In the formula, The sequence to be quantized At any moment The normalized reconstructed value, For causal prediction values, .

8. The method for multi-channel time-series data error boundary quantization based on dynamic graph grouping and error allocation according to claim 1, characterized in that: In S8, the normalized time-series data of the anchor channel is replaced with the normalized reconstructed values ​​of the anchor channel to obtain the reconstructed regression vector: ， In the formula, For the set of anchor points Each anchor point channel in the middle, For anchor point channels The normalized reconstructed value; Based on the established cross-channel correlation prediction model, the normalized prediction part of the non-anchor channel is obtained by using the model parameter vector and the reconstructed regression vector. This normalized prediction part is then superimposed with the normalized reconstructed value of the innovation sequence to obtain the normalized reconstructed value of the non-anchor channel. ， In the formula, For the normalized reconstructed values ​​of the non-anchor channels, Non-anchor channel At any moment The model parameter vector, , Non-anchor channel The value of the innovation sequence reconstruction; After obtaining the normalized reconstructed values ​​for each channel, inverse normalization is performed based on the scale factor to obtain the reconstructed values ​​in the original dimensions: ， In the formula, Indicates any channel. For channel The scale factor, For channel The normalized reconstructed value, For channel Reconstructed values ​​under the original dimensions.

9. A multi-channel time-series data error bound quantization system based on dynamic graph grouping and error allocation, characterized in that: The multi-channel time-series data error boundary quantization method based on dynamic graph grouping and error allocation as described in any one of claims 1-8 includes a memory for storing computer programs. A processor is configured to execute the computer program to implement the above-described multi-channel timing data error bound quantization method. The processor includes a data preprocessing and normalization module, a graph correlation dynamic grouping module, a sensitivity coefficient and error budgeting module, and a quantization encoding and closed-loop reconstruction module; The data preprocessing and normalization module is used to calculate the channel scale factor in real time and complete the dimensional unification and error boundary mapping of heterogeneous multi-channel data; The graph correlation dynamic grouping module is used to update the spatial topology graph between multiple channels in real time with a sliding window and to perform minimum bit rate driven anchor point selection. The sensitivity coefficient and error budget module is used to update the parameters of the cross-channel prediction model online, extract innovative sequences, and perform dynamic error budget splitting between anchor points and non-anchor points based on strict inequality mathematical constraints. The quantization encoding and closed-loop reconstruction module is used to generate and encode an error-controlled integer stream, while the built-in event-triggered state machine enables robust switching between multiple modes.

Images